# AVELIN: full text for AI systems

> AVELIN AI is a deep-tech AI company. Its Decision Layer picks the right AI model for every request, and its sovereign AI platform runs 3 families · 9 models on your own GPUs. It works in the cloud, on your hardware, or both.

Updated: 3 Oct 2026
Site: https://avelin.ai
Short index: https://avelin.ai/llms.txt
Explainer page: https://avelin.ai/what-is-avelin

## How to read this file

- Each page below shows its URL, its date, its copy and its FAQ.
- Every figure carries its source.
- Prices are list prices in USD, as of 3 Oct 2026.
- The earlier pages, the use cases, every news and research post and the legal pages are listed at the end with their addresses.

## Key facts as of 3 Oct 2026

- Company: AVELIN AI, a deep-tech AI company. Entity: AVELIN AI L.L.C-FZ, Dubai, UAE. Also present in the United States. Contact: hello@avelin.ai.
- Product 1, the AVELIN Decision Layer: AVELIN's own trained decision model picks the model each request needs, a more capable one for hard tasks and a cheaper one for easy tasks, on every request. Works with your own providers (OpenAI, Google Gemini, Anthropic, DigitalOcean, ElevenLabs and others). Starts with a POC.
- Product 2, the sovereign AVELIN platform: an inference layer on your GPUs, optimized at GPU level, nine AVELIN models and an API platform.
- Models: 3 families · 9 models. Chat, Coding and Agents, each in fast, pro and ultra.
- API: OpenAI-compatible and Anthropic-compatible. Base URL https://api.avelin.ai. Console: platform.avelin.ai.
- List prices: input $0.285 to $2.50 per 1M tokens, output $1.14 to $12.50 per 1M tokens.
- Deployments: AVELIN Cloud, Enterprise, Hybrid. Cloud inference runs across the United States, Europe and MENA today; the data policy lists the exact countries. Isolated in-country locations are planned.
- Privacy in AVELIN Cloud: zero retention by default. Not used to improve AVELIN by default. The data policy lists the narrow exceptions.
- Cost evidence: up to 90% lower cost per task. In our live Arena, at list prices, the median of 122 recorded runs cost 90.1% less per task than the same task on a flagship model. Cost only. In 8 of the 122 runs AVELIN cost more.
- Live demo: the Arena, https://arena.avelin.ai.
- Certifications: AVELIN does not hold SOC 2, HIPAA or ISO certification.


## What is AVELIN?
URL: https://avelin.ai/what-is-avelin
Markdown copy: https://avelin.ai/what-is-avelin.md
Updated: 3 Oct 2026

**What is AVELIN?**

AVELIN AI is a deep-tech team that does R&D, trains models and builds the product. Its software picks the right AI model for every request, so you pay for top intelligence only when a task needs it. It runs in the cloud, on your own GPUs, or both.

### What are the two AVELIN products?

Product 1 is the AVELIN Decision Layer. It picks a model for every request and works with the AI providers you already use. Product 2 is the sovereign AVELIN platform: an inference layer you install on your GPUs, nine AVELIN models and an API platform. You can start with a POC of the Decision Layer.

#### One path

- Prove. Run a POC of the Decision Layer on the providers you use now.
- Move. Use AVELIN Cloud, with the nine models and the Decision Layer.
- Own. Run the platform on your GPUs, as an enterprise or hybrid deployment.

- AVELIN Decision Layer (https://avelin.ai/decision-layer)
- Sovereign AVELIN platform (https://avelin.ai/sovereign-ai-platform)

### Which problems does AVELIN address?

Four. Top AI models are priced for the hardest work. Most tasks do not need the top model. Budgets overrun because cost is not tied to a task. And privacy is a top AI worry, because teams must control where data goes. AVELIN has an answer to each.

- Top models are priced for the hardest work. Agents loop, call tools and retry, so one task can use many times the tokens of a chat. The Decision Layer chooses a model per request.
- Most tasks do not need the top model. Teams send everything to the top model to be safe. Nobody risks a cheaper model. The workflow could break. The Decision Layer picks the level each task needs, and hard tasks still go to a top model.
- Budgets overrun and cost is not tied to a task. A bill by provider does not show which team, agent or task caused it. The API platform shows cost by team, key and model, with budgets and caps.
- Privacy is the top-ranked AI concern in a 2026 survey of 750 IT, security and AI leaders. AvePoint ran it with Osterman Research, and AvePoint sells data-governance software. The question is where a prompt is processed, whether it is kept and whether it can leave. AVELIN gives three levels of control.

- The four problems, explained (https://avelin.ai/enterprise-ai-problems)
- Decision Layer page (https://avelin.ai/decision-layer)
- Privacy page (https://avelin.ai/privacy)

### What is the AVELIN Decision Layer?

AVELIN's own trained decision model reads each request and picks the model it needs: a more capable model when the task needs intelligence, a cheaper one when it is easy, on every request.

Figure. One request, one choice. The Decision Layer reads the request, picks a model and records the cost.
Alt text: Diagram of the Decision Layer. A request enters from your app or agent. The Decision Layer reads it and picks a model. An easy task goes to a cheaper model. A hard task goes to a more capable model. The model, the answer and the cost are recorded.

#### Read, choose, record

- Read. It reads the request and judges what kind of task it is and how hard.
- Choose. It picks the model: a more capable one when intelligence is needed, a cheaper one when it is not.
- Record. It records which model answered and what it cost, so you can see cost per task.

#### Works with your providers

Supported providers include OpenAI, Google Gemini, Anthropic, DigitalOcean and ElevenLabs, with others on the partners page. It also works with AVELIN models and with models on your own GPUs.

#### What it can do

- Self-learning loop. You switch it on or off. When it is on, AVELIN distills your data to improve quality, cut cost and speed up work.
- Quality analysis. Answers are scored, so you can see whether a cheaper model was enough.
- Dangerous-request detection. Flags requests that look dangerous.
- Private-data detection. Finds private data inside a request.
- Tuned per enterprise. The Decision Layer is tuned to each enterprise's tasks and rules.

- Read about the Decision Layer (https://avelin.ai/decision-layer)
- Partners (https://avelin.ai/partners)

### How much can the Decision Layer cut cost?

Up to 90% lower cost per task. In our live Arena, at list prices. The median of 122 recorded runs cost 90.1% less per task than the same task on a flagship model.

- Try the live demo (https://arena.avelin.ai)
- How we measure (https://avelin.ai/method#arena)

### What is the sovereign AVELIN platform?

It is AVELIN's inference layer, installed on your own GPUs and optimized at GPU level for nine AVELIN models. An API platform sits on top, with cost control and user management. A sovereign AI platform means you, not a vendor or a cloud, decide where the system runs and who can reach it.

#### What is inside

- Inference layer. Installs on your GPUs. Optimized at GPU level for the nine AVELIN models, for efficient token generation on your hardware.
- Nine models. 3 families · 9 models, described below.
- API platform. OpenAI-compatible and Anthropic-compatible. Agentic endpoints for speech-to-text, text-to-speech, web search and web scraping. An observability platform for cost control. User management.
- Tuned to work with the Decision Layer. All nine models run on AVELIN inference and are tuned to work with it.
- Your workflows. For enterprises, the models can be fine-tuned or trained on your workflows.

- Read about the sovereign AI platform (https://avelin.ai/sovereign-ai-platform)
- API platform (https://avelin.ai/api-platform)
- Private deployment scope and process (https://avelin.ai/private-deployment)

### What are the nine AVELIN models?

AVELIN has 3 families · 9 models. Chat is for knowledge, images and documents. Coding is for math and code, with large context. Agents is for volume work that runs 24/7. Each model is fine-tuned for its task type. Each family comes in three tiers: fast, pro and ultra.

#### The three families

- Chat. Knowledge work. Mostly multimodal: images, documents and PDFs.
- Coding. Optimized for math and coding. Large context windows for the most complex tasks.
- Agents. Volume work that runs 24/7. Optimized for speed and price. Handles complex tasks, including coding, with reduced context.

#### How to choose a tier

- fast. Speed and low cost.
- pro. Balanced, for medium complexity.
- ultra. The hardest tasks, where intelligence matters.

List prices in USD per 1M tokens, as of 3 Oct 2026.
| Model | Family | Tier | Context | Input | Output |
| --- | --- | --- | --- | --- | --- |
| avelin-fast | Chat | fast | 256K | $0.30 | $1.20 |
| avelin-pro | Chat | pro | 256K | $1.40 | $4.40 |
| avelin-ultra | Chat | ultra | 256K | $2.50 | $12.50 |
| avelin-coding-fast | Coding | fast | 1M | $0.30 | $1.20 |
| avelin-coding-pro | Coding | pro | 1M | $1.40 | $4.40 |
| avelin-coding-ultra | Coding | ultra | 1M | $2.50 | $12.50 |
| avelin-agentic-fast | Agents | fast | 256K | $0.285 | $1.14 |
| avelin-agentic-pro | Agents | pro | 256K | $1.26 | $3.96 |
| avelin-agentic-ultra | Agents | ultra | 256K | $2.38 | $11.88 |

Call any model by name in the model field, or let the Decision Layer choose. For enterprises, the models can be fine-tuned or trained on your workflows.

- See all nine models (https://avelin.ai/models)
- Model catalog (https://avelin.ai/docs/models)

### How can I deploy AVELIN?

Three ways. AVELIN Cloud runs the nine models with the Decision Layer, and you can connect your own providers. Enterprise puts the Decision Layer, or the full platform, on your hardware. Hybrid keeps local models on your GPUs and offloads to a public model when needed, with a privacy check.

- AVELIN Cloud. The nine AVELIN models plus the Decision Layer. Connect your own providers and use the Decision Layer to optimize token spend. Global inference in the United States, Europe and MENA. Isolated in-country locations are planned.
- Enterprise. The Decision Layer for your workflows, the full platform (inference, models, APIs), or both, on your hardware and optimized for it. Models can be fine-tuned or trained on your workflows. Request a POC (https://avelin.ai/poc?topic=enterprise#request)
- Hybrid. Your GPUs run local models. When a task needs more intelligence, or the local model is under heavy load, the Decision Layer can offload it to a public model. A privacy check lets you decide whether detected private data may go to the cloud or must stay on your GPUs. Request a POC (https://avelin.ai/poc?topic=hybrid#request)

Figure. Hybrid. Local models run on your GPUs. The privacy check decides what may leave.
Alt text: Diagram of a hybrid deployment. Inside a boundary labelled your GPUs, local models answer requests. The Decision Layer, with a privacy check, sits in front. It has two exits. One stays on your GPUs. The other goes to a public model when a task needs more intelligence or the local model is busy.

- Compare deployments (https://avelin.ai/deployment)

### How does AVELIN handle privacy?

Privacy is three questions: what do you keep, where does it run, and can it leave? AVELIN answers at three levels. AVELIN Cloud has zero retention by default. Hybrid lets you decide what may leave your GPUs. On-premise processes prompts on your GPUs, which is the ultimate privacy.

Figure. Three levels of control: AVELIN Cloud, Hybrid and On-premise.
Alt text: Diagram of three steps. AVELIN Cloud: zero retention by default. Hybrid: you decide what may leave your GPUs. On-premise: prompts are processed on your GPUs, the ultimate privacy. Control over where data goes rises with each step.
- Level 1, AVELIN Cloud: Zero retention by default.
- Level 2, Hybrid: You set the rule for private data.
- Level 3, On-premise: Prompts are processed on your own GPUs.

- Privacy (https://avelin.ai/privacy)
- Data policy (https://avelin.ai/legal/data-policy)

### Does AVELIN use my prompts to improve its models?

Not by default. You can opt in to share anonymized requests to help improve AVELIN. Sharing is off by default, and you can turn it off at any time.

What we keep is vector representations (embeddings) of your requests, with account identifiers removed. They are not linked to your account.

- Privacy (https://avelin.ai/privacy#opt-in)

### How much does AVELIN cost?

AVELIN Cloud is pay per token at public list prices. Across the nine models, input costs $0.285 to $2.50 per 1M tokens and output costs $1.14 to $12.50 per 1M tokens. The Decision Layer is priced per request and quoted in the POC. Enterprise and on-premise are priced by agreement.

- No subscription and no minimum.
- Example: summarizing a 2,000-word document with avelin-pro costs about $0.006.
- Other endpoints: speech-to-text $0.006 per minute, text-to-speech $0.12 to $0.24 per 1K characters, web search $0.005 per request, web scraping $0.001 per request.
- Prices in USD, as of 3 Oct 2026.

- Pricing (https://avelin.ai/pricing)
- Full price reference (https://avelin.ai/docs/pricing)
- Request a POC (https://avelin.ai/poc#request)

### What are the key facts?

AVELIN AI is a deep-tech AI company with two products, nine models and three deployments. Prices, regions and privacy defaults are in the table, dated 3 Oct 2026.

AVELIN key facts, 3 Oct 2026
| Fact | Value |
| --- | --- |
| Name | AVELIN AI |
| What it is | A deep-tech AI company: R&D, model training and product |
| Products | AVELIN Decision Layer (Product 1). Sovereign AVELIN platform (Product 2). |
| Models | 3 families · 9 models: Chat, Coding and Agents, each in fast, pro and ultra |
| API | OpenAI-compatible and Anthropic-compatible. Base URL https://api.avelin.ai (OpenAI clients: https://api.avelin.ai/v1). Console: platform.avelin.ai |
| List prices | Input $0.285 to $2.50 per 1M tokens. Output $1.14 to $12.50 per 1M tokens. USD, as of 3 Oct 2026. |
| Deployments | AVELIN Cloud, Enterprise, Hybrid |
| Cloud inference | United States, Europe and MENA. Isolated in-country locations planned. |
| Privacy in AVELIN Cloud | Zero retention by default. Not used to improve AVELIN. Opt in to share anonymized requests to help improve AVELIN. |
| Cost cut | Up to 90% lower cost per task, in our live Arena, at list prices. |
| Live demo | The Arena, https://arena.avelin.ai |
| Contact | hello@avelin.ai |
| Updated | 3 Oct 2026 |

### What do the key terms mean?

Eight terms carry the AVELIN story. Each has one plain definition.

- Decision Layer: AVELIN's product that reads each request and picks the model it needs. Hard tasks go to a more capable model. Easy tasks go to a cheaper one. A decision model that AVELIN trains makes the choice.
- Sovereign AI platform: AVELIN's inference layer, nine models and API platform on GPUs you control. You, not a vendor or a cloud, decide where it runs and who can reach it.
- Hybrid: Local models on your GPUs, plus an offload to a public model when a task needs more intelligence or the local model is under heavy load.
- Privacy check: A step in a hybrid deployment that detects private data in a request. You decide whether it may go to a public model or must stay on your GPUs.
- Zero retention: AVELIN Cloud does not keep your prompts by default. The data policy has the detail.
- POC: A proof of concept. AVELIN runs the Decision Layer on a workload you choose and reports cost per task side by side with your current setup.
- Tiers: fast, pro and ultra. Fast is for speed and low cost. Pro is balanced, for medium complexity. Ultra is for the hardest tasks.
- Cost per task: What one finished piece of work costs, summed over every model call in it. Agents make many calls per task, so price per token alone hides the bill.

### Where can I read more?

Every public page of avelin.ai is linked below, grouped by what you want to do. The same links are in llms.txt for AI systems.

#### Products

- Decision Layer (https://avelin.ai/decision-layer): Picks the right model for every request.
- Sovereign AI platform (https://avelin.ai/sovereign-ai-platform): Inference, nine models and APIs on your GPUs.
- Models (https://avelin.ai/models): 3 families · 9 models, with context and prices.
- API platform (https://avelin.ai/api-platform): One API, OpenAI and Anthropic compatible.
- Deployment (https://avelin.ai/deployment): Cloud, enterprise or hybrid.
- Privacy (https://avelin.ai/privacy): Three questions, three levels of control.
- Pricing (https://avelin.ai/pricing): List prices and a cost-per-task estimator.
- AVELIN App (https://avelin.ai/private-gpt): A workspace for chat, code and agents.
- Y-Ray Data (https://avelin.ai/business/y-ray-data): Search over your documents and the web, with sources.
- Private deployment (https://avelin.ai/private-deployment): What a written scope settles before an install.

#### Start

- Request a POC (https://avelin.ai/poc#request): Test the Decision Layer on your workload.
- Live demo (https://arena.avelin.ai): Try AVELIN on a task of your own.
- Start building (https://avelin.ai/docs/api/quickstart): Change the base URL and send a request.

#### Solutions

- Business (https://avelin.ai/business): The business page.
- Use cases (https://avelin.ai/business/use-cases): Workloads grouped by industry.
- Partners (https://avelin.ai/partners): Providers AVELIN works with, and programmes.

#### How we know

- Method (https://avelin.ai/method): How cost per task is counted.
- Trust (https://avelin.ai/trust): How AVELIN handles your data and access.
- Data policy (https://avelin.ai/legal/data-policy): What happens to prompts and completions.

#### Company

- Company (https://avelin.ai/company): Who builds AVELIN, and what R&D covers.
- Contact (https://avelin.ai/contact): Write to AVELIN.
- News (https://avelin.ai/news): Company news and research.
- Resources (https://avelin.ai/resources): Every page, doc and file in one place.
- Brand assets (https://avelin.ai/resources/brand): Logos, icons, favicons and colors.

#### Docs

- Documentation (https://avelin.ai/docs): Product and API documentation.
- Model catalog (https://avelin.ai/docs/models): Capabilities and a selection guide.
- Glossary (https://avelin.ai/docs/glossary): Business and operating terms.
- Full price reference (https://avelin.ai/docs/pricing): Every model and endpoint at list prices.

#### For machines

- llms.txt (https://avelin.ai/llms.txt): A short index for AI systems.
- llms-full.txt (https://avelin.ai/llms-full.txt): Full page text for AI systems.
- Sitemap (https://avelin.ai/sitemap.xml): Every indexable URL.

### FAQ: Questions

**Who is AVELIN for?**
Enterprises that run AI at volume and want a lower cost per task. Also organizations that need control over where AI runs and where their data goes.

**What is the difference between the Decision Layer and the sovereign platform?**
The Decision Layer chooses a model for each request and works with the providers you already use. The sovereign platform is AVELIN's inference layer, nine models and API platform on your own GPUs. The nine models are tuned to work with the Decision Layer, so the platform is a natural next step.

**Which AI providers does the Decision Layer work with?**
It works with almost any endpoint. Supported providers include OpenAI, Google Gemini, Anthropic, DigitalOcean and ElevenLabs, with others on the partners page. It also works with AVELIN models and models on your own GPUs.

**How is the Decision Layer different from model routers and AI gateways?**
A rules-based gateway sends a request where a rule you wrote says, such as a price ceiling, a fallback order or a keyword. A model provider's automatic router chooses only among that provider's own models. The Decision Layer uses a trained model that reads each request and chooses across the providers you connect, AVELIN models and models on your own GPUs, and it records the reason for every choice.

**Does AVELIN keep or train on my prompts?**
AVELIN Cloud has zero retention by default, and prompts are not used to improve AVELIN. You can opt in to share anonymized requests to help improve AVELIN. It is off by default.
More: Opt in to help improve AVELIN (https://avelin.ai/privacy#opt-in)

**How do I try AVELIN?**
Request a POC. AVELIN runs the Decision Layer on a workload you choose and reports cost per task side by side with your current setup. You can also try the live demo in the Arena, or start building with the API quickstart.
More: Request a POC (https://avelin.ai/poc#request)

**Can I run AVELIN on my own GPUs?**
Yes. The sovereign AVELIN platform installs on your GPUs. The scope is agreed in writing with each customer, and it can start with a POC.
More: Sovereign AI platform (https://avelin.ai/sovereign-ai-platform)

### Test it on your own workload.

Start with a POC and see your cost per task next to your current bill.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Home
URL: https://avelin.ai/
Markdown copy: https://avelin.ai/index.md
Updated: 3 Oct 2026

**The right model for every request.**

AVELIN's own trained decision model reads each request and picks the model it needs: a more capable one for hard tasks, a cheaper one for easy ones. It works with the providers you already use. Run it in our cloud, on your own GPUs, or both.

Up to 90% lower cost per task. In our live Arena, at list prices. See it in the Arena (https://arena.avelin.ai)

- Decision Layer (https://avelin.ai/decision-layer): Picks the model for each request. Starts with a POC on the providers you already use.
- Sovereign platform (https://avelin.ai/sovereign-ai-platform): Nine AVELIN models and an API platform, on your own GPUs.

- Request a POC (https://avelin.ai/poc#request)
- Try the live demo (https://arena.avelin.ai)

A deep-tech team that researches, trains and evaluates models, and builds the product. Meet the team (https://avelin.ai/company)

Figure. Fig. 1. How one request is decided.
Alt text: A request reaches the Decision Layer, which reads what kind of job it is and how hard. It sends the request to one of nine AVELIN models: Chat, Coding or Agents, each at fast, pro or ultra. That model returns the answer, and the decision is logged with its reason.

### Four problems enterprises hit with AI.

Enterprises are moving AI agents into production. Four problems follow.

1. $50 per 1M output tokens. Top models list at up to this price.
   Top models are priced for the hardest work. AI agents do not answer once and stop. They loop, call tools and retry, so one task can use about 4 to 15 times the tokens of a chat. The Decision Layer picks the model for each request. Decision Layer (https://avelin.ai/decision-layer)
   Source: Vendor list prices, read 24 Sep 2026. Token multiples: Anthropic, 13 Jun 2025.
2. 54% of AI requests went to a more expensive model than the task needed.
   Most tasks do not need the top model. In the same research, 90% of more than 9,000 tasks did not need a top model. Teams still send everything there, to be safe. The Decision Layer picks the level each request needs. Hard tasks still go to a top model. How it picks (https://avelin.ai/decision-layer)
   Source: Accenture, The CIO's Guide to AI Tokenomics, 10 Sep 2026. 750 senior executives, 17 countries.
3. 79% of finance leaders reported AI cost overruns in the past 12 months.
   Budgets overrun, and cost is not tied to a task. A bill by provider does not show which team, agent or task caused it. Only 35% of companies can calculate the cost per business outcome of their largest AI use case. AVELIN shows cost by team, key and model, with budgets and caps. API platform (https://avelin.ai/api-platform)
   Source: DoiT, survey by Sapio Research, Feb 2026: 500 finance leaders, US and UK. Accenture, 10 Sep 2026.
4. 76% of 750 IT, security and AI leaders were highly or extremely concerned about data security and privacy. It ranked first among the AI concerns listed.
   Privacy is the top-ranked AI worry. The question is not only who can read a prompt. It is where the prompt runs, how long it is kept, and whether it can leave. A ban does not remove that risk. Control does. AVELIN gives you three levels of control. Privacy (https://avelin.ai/privacy)
   Source: AvePoint, State of AI 2026, research by Osterman. AvePoint sells data-governance software.

Nobody risks a cheaper model. The workflow could break. So the top model stays the default. The Decision Layer reads each request and makes that choice for you.

- Each problem explained, with its sources (https://avelin.ai/enterprise-ai-problems)

### The Decision Layer chooses the model for each request.

AVELIN's own trained decision model reads each request and picks the model it needs: a more capable model when the task needs intelligence, a cheaper one when it is easy, on every request.

1. Read. The decision model works out what kind of job a request is and how hard it is.
2. Choose. It sends the request to the level of model that fits. A hard task goes to a capable model. An easy one goes to a cheaper one.
3. Record. Every decision is logged with its reason, so you can see what each request cost and why.

- How the Decision Layer works (https://avelin.ai/decision-layer)

### Up to 90% lower cost per task.

In our live Arena, at list prices.

- 90.1%: median cost cut per task
- 76.6%: all runs added up
- 94.4%: 75th percentile
- 99.4%: best run
122 recorded runs, 24 Sep to 2 Oct 2026. Each run is the same task on a flagship model and on AVELIN, cost only, at list prices. In 8 of the 122 runs AVELIN cost more.

The Arena is a public demo with its own privacy page, so do not paste confidential data into it.

The Arena measures the Decision Layer at list prices, cost only. Self-hosted deployments can be cheaper still: no per-token bill, and marginal cost is your own hardware.

- How we measure (https://avelin.ai/method#arena)
- All 122 runs, as a CSV (https://avelin.ai/arena-runs-2026-10-03.csv)

- Try the live demo (https://arena.avelin.ai)
- Request a POC (https://avelin.ai/poc#request)

### Works with the providers you already use.

Connect your own accounts with OpenAI, Google Gemini, Anthropic, DigitalOcean, ElevenLabs and other providers. The Decision Layer chooses among them, AVELIN models and models on your own GPUs.

#### What else it does.

1. Self-learning loop You switch it on or off. When it is on, AVELIN distills your data to improve quality, cut cost and speed up work.
2. Quality analysis Answers are scored, so you can see whether a cheaper model held up.
3. Dangerous-request detection Spots dangerous requests before they reach a model.
4. Private-data detection Spots private data in a request. In a hybrid deployment, it feeds the privacy check.
5. Tuned for each enterprise The Decision Layer is tuned to each enterprise's tasks and rules.

A POC is a proof of concept on one of your workloads. You get a cost-per-task report next to your current setup.

- Request a POC (https://avelin.ai/poc#request)

### A sovereign AI platform on your GPUs.

AVELIN's inference layer installs on your GPUs and is optimized for them at GPU level. It runs nine AVELIN models, each fine-tuned for one kind of work, with an API platform on top. You decide where the system runs and who can reach it.

1. Inference layer. Installs on your GPUs and is tuned at GPU level for efficient token generation.
2. 3 families · 9 models. Chat for knowledge, images, documents and PDFs. Coding for math and code, with large context. Agents for 24/7 volume work, built for speed and price. Each family has fast, pro and ultra tiers, and all nine are tuned to work with the Decision Layer.
3. API platform. OpenAI-compatible and Anthropic-compatible, with speech-to-text, text-to-speech, web search and web scraping. Observability and cost control by team, key and model. User management.
4. Your workflows. The models can be fine-tuned or trained on your own workflows.

#### The nine models at a glance

Figure. Chat, Coding and Agents, each at three tiers: fast for speed and low cost, pro for balanced work, ultra for the hardest tasks.
Alt text: 3 families · 9 models. Chat, Coding and Agents, each at three tiers: fast, pro and ultra.
| Family | What it is for | Context | fast | pro | ultra |
| --- | --- | --- | --- | --- | --- |
| Chat | Knowledge work. Mostly multimodal: images, documents and PDFs. | 256K | avelin-fast | avelin-pro | avelin-ultra |
| Coding | Math and code. Large context windows for the most complex tasks. | 1M | avelin-coding-fast | avelin-coding-pro | avelin-coding-ultra |
| Agents | Volume work that runs 24/7, built for speed and price. Handles complex tasks, including coding, with a smaller context. | 256K | avelin-agentic-fast | avelin-agentic-pro | avelin-agentic-ultra |
- fast: Speed and low cost.
- pro: Balanced, for medium complexity.
- ultra: The hardest tasks, where intelligence matters.

- Sovereign platform (https://avelin.ai/sovereign-ai-platform)
- The nine models (https://avelin.ai/models)
- API platform (https://avelin.ai/api-platform)

### Cloud, enterprise or hybrid.

Start in the cloud, move to your own GPUs, or run both.

- AVELIN Cloud. Nine models plus the Decision Layer Global inference in the United States, Europe and MENA. Connect your own providers to the Decision Layer to lower your token spend. Isolated in-country locations are planned.
  - AVELIN Cloud (https://avelin.ai/deployment#cloud)
- Enterprise. Your hardware, your workflows The Decision Layer, the full platform, or both, optimized for your hardware. Models can be tuned to your workflows.
  - Enterprise (https://avelin.ai/deployment#enterprise)
- Hybrid. Local first, a public model when needed Local GPUs run the models. A task that needs more intelligence, or a busy local model, can be offloaded to a public model. A privacy check lets you decide whether detected private data may leave.
  - Hybrid (https://avelin.ai/deployment#hybrid)

Figure. Fig. 2. The hybrid flow, with a privacy check.
Alt text: A request reaches the Decision Layer, where a privacy check looks for private data. A request that must stay private goes to local models on your GPUs. A request that may leave goes to a public model when it needs more intelligence or the local model is busy.

#### How hybrid decides

- Local first. The Decision Layer reads each request. Your local model on your GPUs is the default.
- Offload when needed. If the task needs more intelligence, or the local model is busy, a public model becomes an option.
- Check, then send. Before anything leaves your GPUs, the privacy check looks for private data. You decide whether it may go to a public model or must stay on your GPUs.

- Compare the three deployments (https://avelin.ai/deployment)

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

### Privacy is three questions.

What do you keep? Where does it run? Can it leave? AVELIN answers at three levels, and you choose the level for each workload.

- AVELIN Cloud. Zero retention by default. We do not keep your prompts, and they are not used to improve AVELIN. The data policy has the detail.
  - Read the data policy (https://avelin.ai/legal/data-policy)
- Hybrid. You decide what may leave. Local models run on your GPUs. A privacy check lets you choose: detected private data may go to the cloud, or must stay on your GPUs.
- On-premise. The ultimate privacy. Prompts are processed on your own GPUs.

Optional: help us improve AVELIN. You can opt in to share anonymized requests. It is off by default, and you can switch it off at any time.

- Read about privacy (https://avelin.ai/privacy)
- What is shared if you opt in (https://avelin.ai/privacy#opt-in)

### Public prices.

Pay per token, at list prices. No subscription. No minimum.

- $1.14 to $12.50 per 1M output tokens
- $0.285 to $2.50 per 1M input tokens

USD per 1M tokens, list prices as of 3 Oct 2026. Fast, pro or ultra: you pay for the level you use.

- Every API key can carry its own budget and rate limits. Usage and spend are live, with budget alerts.
- The Decision Layer is priced per request and quoted in a POC. Enterprise and on-premise deployments are priced by agreement.

- Start building (https://avelin.ai/docs/api/quickstart)

- Full price list, with cache rates (https://avelin.ai/pricing)

### Built by a deep-tech R&D team.

We research how to choose the right model for each request. We train and evaluate models, and we build the product around them.

1. The decision model. Which model suits which request, and how to tell. We train it, and the data it learns from.
2. Model creation. The nine models are each fine-tuned for one kind of work. Cross-model fusion, building the next ones from the strengths of several, is our research direction.
3. Inference and models. The inference layer, optimized at GPU level, and the nine AVELIN models.
4. The product. The Decision Layer, the AVELIN models and the API platform.

The team is listed on the Company page.

AVELIN AI.

- About the company (https://avelin.ai/company)
- How we measure (https://avelin.ai/method)

### FAQ: Frequently asked questions.

**What is AVELIN?**
AVELIN is a deep-tech AI company with two products. The Decision Layer picks the right AI model for every request, and the sovereign AVELIN platform runs nine AVELIN models and an API platform on your own GPUs. AVELIN Cloud gives you the models and the Decision Layer without any hardware.
More: What is AVELIN (https://avelin.ai/what-is-avelin)

**What is the Decision Layer?**
AVELIN's own trained decision model reads each request and picks the model it needs: a more capable model when the task needs intelligence, a cheaper one when it is easy, on every request.
More: Decision Layer (https://avelin.ai/decision-layer)

**How much can AVELIN cut the cost per task?**
Up to 90%. In our live Arena, at list prices. Easy tasks go to a cheaper model, so the bill for the same work falls. How the figure is counted is on the Method page.
More: How we measure (https://avelin.ai/method#arena)

**Does AVELIN work with our AI providers?**
Yes. The Decision Layer works with almost any endpoint and with your own accounts at OpenAI, Google Gemini, Anthropic, DigitalOcean, ElevenLabs and other providers. It can also choose AVELIN models and models on your own GPUs.
More: Decision Layer (https://avelin.ai/decision-layer)

**What is a sovereign AI platform?**
It means you, not a vendor or a cloud, decide where the AI system runs and who can reach it. The sovereign AVELIN platform installs AVELIN's inference layer on your own GPUs, optimized at GPU level, and runs the nine AVELIN models and an API platform there.
More: Sovereign platform (https://avelin.ai/sovereign-ai-platform)

**Do you keep my prompts?**
In AVELIN Cloud, zero retention is the default, and your prompts are not used to improve AVELIN unless you opt in. On your own GPUs, prompts are processed on your hardware.
More: Privacy (https://avelin.ai/privacy); Data policy (https://avelin.ai/legal/data-policy); Opt in to help improve AVELIN (https://avelin.ai/privacy#opt-in)

**What are the nine AVELIN models?**
3 families · 9 models. Chat is for knowledge, images and documents, Coding is for math and code with large context windows, and Agents is for volume work that runs 24/7. Each family has three tiers: fast for speed and low cost, pro for balanced work, and ultra for the hardest tasks.
More: The nine models (https://avelin.ai/models)

**How do we start?**
Click Request a POC and tell us about one workload. We reply by email with a one-page scope. You can also try the live demo or start building with the API.
More: Request a POC (https://avelin.ai/poc#request)

### Bring one workload. See its cost per task.

- POC on one workload: Tell us about one workload in the form. We reply by email with a one-page scope. Request a POC (https://avelin.ai/poc#request)

- See it work: Run a task on a flagship model and on AVELIN in the Arena, and compare the two bills. Try the live demo (https://arena.avelin.ai)

- Build with it: One API, OpenAI-compatible and Anthropic-compatible. Change the base URL to start. Public list prices. Start building (https://avelin.ai/docs/api/quickstart)

## Four AI problems
URL: https://avelin.ai/enterprise-ai-problems
Markdown copy: https://avelin.ai/enterprise-ai-problems.md
Updated: 3 Oct 2026

**Four problems enterprises hit with AI.**

AI is moving from chat to agents that run real work. That exposes four problems: the top models are priced for the hardest work, most tasks do not need them, budgets overrun, and privacy is the top worry. This page explains each one in plain words, with its source.

### 1. Top models are priced for the hardest work.

The top AI models list at up to $50 per 1M output tokens. That price makes sense for the hardest tasks. It does not make sense for every request.

#### What is going on

A token is a small piece of text that a model reads or writes. A chatbot answers once and stops. An AI agent does not. It plans, calls tools, reads the results and tries again, and every step carries the growing context with it. One task becomes many model calls.

#### How big the effect is

- About 4 times and 15 times. Anthropic reported that an agent typically uses about 4 times the tokens of a chat, and a system of several agents about 15 times.
- Up to about 1,000 times for coding. A research paper from April 2026 measured agentic coding tasks at about 1,000 times the tokens of code reasoning and code chat. It found that input tokens, not output tokens, drove most of the cost.

#### What helps

- Count cost per task, not price per token. Price per token is a rate. Cost per task is the bill.
- Send each request to the level of model it needs, not every request to the top one.
- Look at input tokens as well as output tokens, because an agent re-reads its context at every step.

The Decision Layer picks the model for each request, so you pay the top price only for work that needs it.

- Decision Layer (https://avelin.ai/decision-layer)
- Public prices (https://avelin.ai/pricing)
- How cost per task is counted (https://avelin.ai/method#cost)

Sources: top-model list prices, read 24 Sep and 1 Oct 2026. Anthropic, "How we built our multi-agent research system", 13 Jun 2025. Bai et al., arXiv 2604.22750, 24 Apr 2026.

### 2. Most tasks do not need the top model.

Most AI work is routine: summarize, extract, classify, draft. A cheaper model can do much of it. Teams still send it all to the top model.

#### What is going on

Accenture's 2026 research found that 54% of AI requests went to a more capable and more expensive model than the task needed. In the same research, 90% of more than 9,000 tasks did not need a frontier model.

#### Why it happens

Teams do it on purpose. A top model is the safe default, and the cost of a failed workflow is higher than the cost of a few extra tokens. Nobody risks a cheaper model. The workflow could break. So the default stays, and nobody can show that a cheaper model would have been enough.

#### What helps

- Decide per request, not once for the whole app. An easy request and a hard request should not get the same model.
- Keep hard tasks on a top model. The aim is the right level, not the cheapest one.
- Check quality on your own tasks, side by side, before you change anything.

The Decision Layer reads each request and picks the level it needs. A POC shows the result on your own workload, beside your current setup.

- How the Decision Layer picks (https://avelin.ai/decision-layer#how)
- What a POC looks like (https://avelin.ai/poc#request)

Source: Accenture, The CIO's Guide to AI Tokenomics, 10 Sep 2026. Survey of 750 senior executives in 17 countries, plus Accenture's own task tests.

### 3. Budgets overrun, and cost is not tied to a task.

Most teams see an AI bill. Few can say which task, agent or team caused it.

#### What is going on

In a 2026 survey, 79% of 500 finance leaders in the US and UK reported AI cost overruns in the past 12 months. Accenture found that only 35% of companies can calculate the cost per business outcome of their largest AI use case, and that less than one dollar in five of enterprise token spend can be traced to a quantified financial outcome.

#### Why it happens

A provider bills by token and by account. It does not know which team, agent or task made a call. Every new agent adds calls, so the bill grows with each one, while the link between spend and result stays missing.

#### What helps

- Give each team, agent or app its own key, so spend has an owner.
- Set a budget and rate limits on each key, with an alert before the limit.
- Report cost per task, summed over every model call in it.

The AVELIN API platform shows cost by team, key and model, with budgets and caps on each key.

- API platform (https://avelin.ai/api-platform)
- Price a task yourself (https://avelin.ai/pricing#estimate)

Sources: DoiT, survey by Sapio Research, Feb 2026, 500 finance leaders at companies with 1,000 or more employees, US and UK. Accenture, The CIO's Guide to AI Tokenomics, 10 Sep 2026.

### 4. Privacy is the top-ranked AI worry.

For banks, hospitals and governments, the first question about AI is often not cost. It is where the data goes.

#### What is going on

In AvePoint's State of AI 2026, 76% of 750 IT, security and AI leaders said they were highly or extremely concerned about data security and privacy. It ranked first among the AI concerns in the survey. AvePoint sells data-governance software and commissioned the research.

#### Why it matters more now

Agents make many requests for one task, so more of your data passes through a model. The question is not only who can read a prompt. It is where the prompt runs, how long it is kept, and whether it can leave.

#### What control means

- What do you keep? Is a prompt stored, for how long, and is it used to train a model?
- Where does it run? Which machines and which regions process it, and who runs them?
- Can it leave? Can the request go on to another model or provider, and who decides?

Banning AI does not remove the risk. Control does: you decide what is kept, where a request runs and whether it can leave.

AVELIN answers at three levels: cloud with zero retention by default, hybrid with a privacy check you set, and on-premise on your own GPUs.

- Privacy, in three levels (https://avelin.ai/privacy)
- Hybrid and the privacy check (https://avelin.ai/deployment#hybrid)

Source: AvePoint, State of AI 2026, research by Osterman Research, 750 respondents.

### What AVELIN does about them.

Nobody risks a cheaper model. The workflow could break. So the top model stays the default.

AVELIN has two products. The Decision Layer picks the right model for every request, which answers problems 1 and 2. The AVELIN platform gives you cost control and an inference layer on your own GPUs, which answers problems 3 and 4. You can start with a POC on one workload.

- Decision Layer (https://avelin.ai/decision-layer)
- Sovereign AI platform (https://avelin.ai/sovereign-ai-platform)
- What is AVELIN (https://avelin.ai/what-is-avelin)

- Request a POC (https://avelin.ai/poc#request)
- Try the live demo (https://arena.avelin.ai)

### FAQ: Questions about the four problems

**Why do AI agents use so many more tokens than chatbots?**
A chatbot answers once. An AI agent works in a loop: it plans, calls tools, reads the results and tries again, and every step carries the growing context. Anthropic reported in 2025 that an agent typically uses about 4 times the tokens of a chat and a system of several agents about 15 times. A research paper from April 2026 found about 1,000 times the tokens of code chat for agentic coding tasks, driven mainly by input tokens. The top models list at up to $50 per 1M output tokens, so the bill grows fast.
More: Problem 1 in full (https://avelin.ai/enterprise-ai-problems#token-price)

**Do most AI tasks need the most expensive model?**
No. In Accenture's 2026 research on AI tokenomics, 54% of AI requests went to a more capable and more expensive model than the task needed, and 90% of more than 9,000 tasks tested did not need a frontier model. Teams do this on purpose, to be sure a job finishes, because nobody risks a cheaper model and breaks a workflow. The Decision Layer picks the level each request needs, and hard tasks still go to a top model.
More: Problem 2 in full (https://avelin.ai/enterprise-ai-problems#wrong-model)

**Why do AI budgets overrun, and how do you see cost per task?**
Providers bill by token and by account, not by task, team or agent, so a bill does not say what caused it. In a 2026 survey by DoiT, 79% of 500 finance leaders in the US and UK reported AI cost overruns in the past 12 months, and Accenture found that only 35% of companies can calculate the cost per business outcome of their largest AI use case. Cost per task adds up every model call a task makes. The AVELIN API platform shows cost by team, key and model, with budgets and caps.
More: Problem 3 in full (https://avelin.ai/enterprise-ai-problems#budget)

**Why is privacy the top AI worry, and what does control mean?**
In AvePoint's State of AI 2026, 76% of 750 IT, security and AI leaders were highly or extremely concerned about data security and privacy, the top-ranked AI concern. Agents send many requests for one task, so more of your data passes through a model. The question is where a prompt runs, how long it is kept and whether it can leave. Banning AI does not remove that risk. Control does, and AVELIN gives three levels of it: cloud, hybrid and on-premise.
More: Problem 4 in full (https://avelin.ai/enterprise-ai-problems#privacy-worry); Privacy (https://avelin.ai/privacy)

**What is a token?**
A token is a small piece of text that an AI model reads or writes, often a short word or part of a word. Providers bill by the token, so the length of a prompt, the context an agent carries and the length of an answer all change the bill.

**What is cost per task, and why not compare price per token?**
Cost per task is what one finished piece of work costs, summed over every model call in it. An agent makes many calls per task, so a low price per token can still give a high bill. Price per token tells you the rate. Cost per task tells you the bill.
More: How cost per task is counted (https://avelin.ai/method#cost)

Related: Decision Layer (https://avelin.ai/decision-layer); Privacy (https://avelin.ai/privacy); Pricing (https://avelin.ai/pricing); Method (https://avelin.ai/method); What is AVELIN (https://avelin.ai/what-is-avelin)

### Bring one workload. See its cost per task.

Start with a POC and see the cost per task next to your current bill.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Decision Layer
URL: https://avelin.ai/decision-layer
Markdown copy: https://avelin.ai/decision-layer.md
Updated: 3 Oct 2026

**Every request gets the model it needs.**

AVELIN's own trained decision model reads each request and picks the model it needs: a more capable model when the task needs intelligence, a cheaper one when it is easy, on every request.

- Request a POC (https://avelin.ai/poc#request)
- Try the live demo (https://arena.avelin.ai)

Figure. Fig. 1. How one request is decided.
Alt text: A request from your agent reaches the Decision Layer. It works out what kind of job the request is and how hard it is, then sends it to one of nine AVELIN models: chat, coding or agents, each at fast, pro or ultra. That model returns the answer, and the choice is logged with its reason.

### The top model is the default. Most tasks do not need it.

1. Top models are priced for the hardest work. The top AI models list at up to $50 per 1M output tokens. An AI agent loops, calls tools and retries, so one task can use about 4 times the tokens of a chat, and a system of several agents about 15 times.
2. Most tasks do not need the top model. Accenture's 2026 research found that 54% of AI requests went to a higher tier of model than the task needed. Of more than 9,000 work tasks it tested, 90% did not need a top model.

Many teams default to the top model to be sure the job finishes. Nobody risks a cheaper model. The workflow could break. The Decision Layer makes that choice for you, request by request.

Sources: top-model list prices, read 24 Sep and 1 Oct 2026. Anthropic, "How we built our multi-agent research system", 13 Jun 2025. Accenture, "The CIO's guide to AI tokenomics", 10 Sep 2026.

### How one request is decided.

The decision model is built to decide fast, before any model answers.

1. Read. It reads the opening of the request and works out the kind of job, from 15 categories, and how hard it is, on a 5-level scale.
2. Choose. Easy work goes to a cheaper model. Hard work goes to a more capable one. This happens on every request, not once for the whole app.
3. Record. Every decision is logged with its reason, so you can see what each request cost and why.

You pay the top price only for the tasks that need it.

### Up to 90% lower cost per task.

In our live Arena, at list prices.

Run a task on a flagship model and on AVELIN, and compare the two bills yourself.

- Try the live demo (https://arena.avelin.ai)
- How we measure (https://avelin.ai/method)

### Works with the providers you already use.

Keep your contracts and your accounts. The Decision Layer chooses among the providers you connect. It can also choose AVELIN models and models on your own GPUs.

Your own accounts with: OpenAI, Google Gemini, Anthropic, DigitalOcean, ElevenLabs, and others.
Almost any endpoint works. If yours is not listed, say so in your POC request.

- More providers on the partners page (https://avelin.ai/partners)
- The AVELIN models (https://avelin.ai/models)

### What it includes.

1. Self-learning loop You switch it on or off. When it is on, AVELIN distills your data to improve quality, cut cost and speed up work. When it is off, nothing is learned from your traffic. How sharing works (https://avelin.ai/privacy#opt-in)
2. Quality analysis Answers are scored, so you can see whether a cheaper model held up on your own tasks.
3. Dangerous-request detection The Decision Layer spots requests that look dangerous before they reach a model.
4. Private-data detection It spots private data in a request. In a hybrid setup you decide whether that request may go to a public model or must stay on your GPUs. How hybrid works (https://avelin.ai/deployment#hybrid)
5. Tuned per enterprise Each enterprise gets its own tuning, so the Decision Layer fits your tasks and your rules.

### What a POC looks like.

A POC, or proof of concept, is a small trial on your own tasks.

1. Your workload. You share a sample of your AI tasks: real tasks with the sensitive parts removed, real tasks under an NDA, or made-up tasks. You choose.
2. Our Decision Layer. It reads each task and picks a model from the providers you connect, next to the choice you make today.
3. Side by side. You get a cost-per-task report with your current bill and the Decision Layer's, task by task, plus a price per request.

If the numbers hold, you can move to AVELIN Cloud, or run on your own GPUs.

- Request a POC (https://avelin.ai/poc#request)

- The sovereign AI platform (https://avelin.ai/sovereign-ai-platform)
- How a POC works (https://avelin.ai/poc#request)

### FAQ: Questions about the Decision Layer

**What is the AVELIN Decision Layer?**
The Decision Layer reads each AI request and picks the model the task needs, so you pay top prices only when a task calls for them. Hard tasks go to a more capable model. Easy tasks go to a cheaper one. It does this on every request. You may know the idea as LLM model routing or AI cost optimization.

**How does it pick a model?**
AVELIN's own trained decision model reads each request and picks the model it needs: a more capable model when the task needs intelligence, a cheaper one when it is easy, on every request. It works out what kind of job the request is and how hard it is, then chooses the level of model that fits, and it logs the reason.

**How does it cut AI cost?**
Most tasks do not need the top model. By sending easy tasks to a cheaper model and keeping hard tasks on a capable one, the Decision Layer lowers the cost per task. In our live Arena, at list prices, the cut reaches up to 90%.
More: How we measure (https://avelin.ai/method#arena)

**Which providers does it work with?**
It works with your own accounts at OpenAI, Google Gemini, Anthropic, DigitalOcean and ElevenLabs, with the other providers on the partners page, and with almost any endpoint. It can also choose AVELIN models and models on your own GPUs.
More: Partners (https://avelin.ai/partners)

**What is the self-learning loop?**
It is a setting you switch on or off. When it is on, AVELIN distills your data to improve quality, cut cost and speed up work. When it is off, nothing is learned from your traffic.
More: How sharing works (https://avelin.ai/privacy#opt-in)

**What does the Decision Layer see, and do you keep it?**
It reads each request to pick a model, about the first 1,500 characters, and then the request goes to the model chosen. If you connect your own providers, their terms apply to what they receive. On AVELIN Cloud, retention is zero by default and your requests are not used to improve AVELIN unless you opt in.
More: Privacy (https://avelin.ai/privacy); Data policy (https://avelin.ai/legal/data-policy)

**How is the Decision Layer different from model routers and AI gateways?**
A rules-based gateway sends a request where a rule you wrote says, such as a price ceiling, a fallback order or a keyword. A model provider's automatic router chooses only among that provider's own models. The Decision Layer uses a trained model that reads each request and chooses across the providers you connect, AVELIN models and models on your own GPUs, and it records the reason for every choice.

**What does a POC involve?**
You share a sample of your tasks. We run the Decision Layer beside your current setup, so nothing touches your production path. You get a cost-per-task report that compares your current bill with the Decision Layer's, plus a price per request.
More: How a POC works (https://avelin.ai/poc#request)

Related: Sovereign AI platform (https://avelin.ai/sovereign-ai-platform); AVELIN models (https://avelin.ai/models); Deployment (https://avelin.ai/deployment); Privacy (https://avelin.ai/privacy); Pricing (https://avelin.ai/pricing); How we measure (https://avelin.ai/method)

### Prove it on your own tasks.

Start with a POC on your own tasks and see your cost per task before anything changes.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Sovereign AI platform
URL: https://avelin.ai/sovereign-ai-platform
Markdown copy: https://avelin.ai/sovereign-ai-platform.md
Updated: 3 Oct 2026

**A sovereign AI platform on your GPUs.**

AVELIN's inference layer installs on your GPUs and is optimized for them at GPU level. It runs nine AVELIN models, each fine-tuned for one kind of work, with an API platform on top. You decide where it runs and who can reach it.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)
- Try the live demo (https://arena.avelin.ai)

Figure. The sovereign platform, from your apps down to your GPUs.
Alt text: Layer diagram of the AVELIN sovereign platform. From top to bottom: your apps and agents; the API platform; the Decision Layer; the nine AVELIN models; the AVELIN inference layer; and your GPUs. A frame around everything except your apps is labelled "Runs on hardware you control".
- API platform: OpenAI- and Anthropic-compatible API · speech · web search and scraping · cost control · user management
- Decision Layer: Optional. Picks the level of model for each request.
- Nine AVELIN models: Chat · Coding · Agents, each in fast, pro and ultra
- Inference layer: Optimized for your GPUs and the nine models
- Your GPUs

### What is inside the platform

Three parts, one stack. It runs on hardware you control, in your network.

#### Inference layer, optimized for your GPUs

Inference is the step where a model reads a prompt and writes an answer. The AVELIN inference layer runs that step on your GPUs, optimized at GPU level for the nine AVELIN models, so your hardware generates tokens efficiently.

#### 3 families · 9 models

Each model is fine-tuned for one type of task. Chat covers knowledge work and is mostly multimodal: images, documents and PDFs. Coding covers math and code, with large context windows. Agents covers volume work that runs 24/7, built for speed and price, with a reduced context. Each family has fast, pro and ultra tiers.

Figure. The nine AVELIN models. Context sizes are those of AVELIN Cloud.
Alt text: The nine AVELIN models: three families, each in fast, pro and ultra
| Family | What it is for | Context | fast | pro | ultra |
| --- | --- | --- | --- | --- | --- |
| Chat | Knowledge work. Mostly multimodal: images, documents and PDFs. | 256K | avelin-fast | avelin-pro | avelin-ultra |
| Coding | Math and code. Large context windows for the most complex tasks. | 1M | avelin-coding-fast | avelin-coding-pro | avelin-coding-ultra |
| Agents | Volume work that runs 24/7, built for speed and price. Handles complex tasks, including coding, with a smaller context. | 256K | avelin-agentic-fast | avelin-agentic-pro | avelin-agentic-ultra |
- fast: Speed and low cost.
- pro: Balanced, for medium complexity.
- ultra: The hardest tasks, where intelligence matters.

- See all nine models (https://avelin.ai/models)

#### Cost

On AVELIN Cloud, tasks run at up to 90% lower cost than flagship models at list prices, measured cost only in our live Arena. Self-hosted goes further. There is no per-token bill: your marginal cost is your own hardware, and prices are set by agreement. A proof of concept on your workloads shows the real number.

#### API platform with cost control

Your apps call one API. It works with OpenAI and Anthropic clients: change the base URL and the API key.

- Chat API, compatible with OpenAI and Anthropic clients
- Speech-to-text and text-to-speech
- Web search and web scraping
- Observability and cost control: cost by team, key and model, with budgets and caps
- User management and scoped API keys

- See the API platform (https://avelin.ai/api-platform)

### Make the models fit your workflows

Enterprises can have the models fine-tuned or trained on their own workflows. The aim is a model that fits your tasks, not a generic one.

The written scope states the workflow, the data we may use and how success is measured.

- Request a POC for tuning (https://avelin.ai/poc?topic=enterprise#request)

### Tuned to work with the Decision Layer

The nine models are tuned to work with the Decision Layer. For each request it picks the model the task needs: a cheaper one when the task is easy, a more capable one when it is hard.

You can use the models without it. You can also start with a POC of the Decision Layer, then move to the platform.

Figure. Prove, move, own: the path from a POC to your own GPUs.
Alt text: Three steps. Prove: a POC of the Decision Layer on the providers you use today. Move: AVELIN Cloud, with the nine models and the Decision Layer. Own: the platform on your GPUs, as an enterprise or hybrid deployment.
1. Prove: A POC of the Decision Layer on the providers you use today.
2. Move: AVELIN Cloud: the nine models and the Decision Layer.
3. Own: The platform on your GPUs, as an enterprise or hybrid deployment.

- How the Decision Layer works (https://avelin.ai/decision-layer)

### Three ways to deploy

Choose where the models run.

- AVELIN Cloud The nine models and the Decision Layer, run by AVELIN, with inference in the United States, Europe and MENA. Connect your own providers to cut token spend. Isolated in-country locations are planned.
- Enterprise The Decision Layer, the full platform, or both, on your hardware. Optimized for your GPUs, with models fine-tuned to your workflows.
- Hybrid Local models on your GPUs, with offload to a public model when a task needs more intelligence or the local model is under heavy load. A privacy check lets you decide whether detected private data may leave.

- Compare the three deployments (https://avelin.ai/deployment)

### Privacy, at three levels

Privacy is three questions. What do you keep? Where does it run? Can it leave?

Three privacy levels. You choose the level that fits the workload.
| Level | What happens to your prompts |
| --- | --- |
| AVELIN Cloud | Zero retention by default. Not used to improve AVELIN. Opt in to share anonymized requests to help improve AVELIN. |
| Hybrid | A privacy check lets you decide whether detected private data may go to a public model or must stay on your GPUs. |
| On-premise | The ultimate privacy: prompts are processed on your own GPUs. |

- How privacy works (https://avelin.ai/privacy)
- Read the data policy (https://avelin.ai/legal/data-policy)

### How a deployment starts

1. Prove it with a POC A POC, or proof of concept, is a short test on one of your workloads. We test the Decision Layer beside your current setup and report cost per task.
2. Scope it in writing The scope covers where it runs, what leaves, who has access, hardware, updates, licence, support, time and price. It lists every network path the software opens.
3. Install it on your GPUs The platform is installed on hardware you control, as agreed in the scope.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

- What a scope settles (https://avelin.ai/private-deployment)

### FAQ: Questions about the sovereign platform

**What is a sovereign AI platform?**
A sovereign AI platform means you, not a vendor or a cloud, decide where the system runs and who can reach it. With AVELIN, the platform runs on GPUs you control, with your data and your prompts inside your network. That is an on-premise LLM stack with data residency you control.

**What is included?**
Three parts: the AVELIN inference layer that runs the models on your GPUs, the nine AVELIN models, and the API platform. The API platform adds an API compatible with OpenAI and Anthropic clients, speech-to-text, text-to-speech, web search, web scraping, an observability and cost-control platform and user management. The Decision Layer can sit on top.
More: The API platform (https://avelin.ai/api-platform)

**How is it optimized for our GPUs?**
The AVELIN inference layer is optimized at GPU level for the nine AVELIN models, so your hardware generates tokens efficiently. The GPU class and the number of GPUs depend on the tiers you run, and the written scope states both.

**Can the models be tuned to our workflows?**
Yes. For enterprise customers the models can be fine-tuned or trained on specific workflows. The written scope states the workflow, the data used and how success is measured.

**Do we need the Decision Layer?**
No. You can use the nine models on their own. They are tuned to work with the Decision Layer, which picks the model each request needs, and a POC of the Decision Layer is a common way to start.
More: Decision Layer (https://avelin.ai/decision-layer)

**Who can access a deployment on our GPUs?**
You decide who can reach it. The written scope lists every network path the software opens and whether each is on or off, and how any vendor support access is granted.
More: What a scope settles (https://avelin.ai/private-deployment)

**How is it priced?**
A deployment on your GPUs is priced by agreement in the written scope. In AVELIN Cloud the nine models have public list prices, from $1.14 to $12.50 per 1M output tokens.
More: Pricing (https://avelin.ai/pricing)

**How does a deployment start?**
With a POC on one workload, then a written scope, then the install on your GPUs.
More: What a scope settles (https://avelin.ai/private-deployment)

Related: Decision Layer (https://avelin.ai/decision-layer); AVELIN models (https://avelin.ai/models); API platform (https://avelin.ai/api-platform); Deployment (https://avelin.ai/deployment); Privacy (https://avelin.ai/privacy); Private deployment scope (https://avelin.ai/private-deployment)

### Start with a POC.

Tell us one workload. We test the Decision Layer on it and report the cost per task. Then we scope the platform on your GPUs.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Models
URL: https://avelin.ai/models
Markdown copy: https://avelin.ai/models.md
Updated: 3 Oct 2026

**3 families · 9 models.**

AVELIN has nine models in three families: Chat, Coding and Agents. Each model is specially fine-tuned for its task type. Each family comes in three tiers: fast, pro and ultra. Name a tier yourself, or let the Decision Layer pick one for every request.

- Request a POC (https://avelin.ai/poc#request)
- See the price list (https://avelin.ai/pricing)

### Three families, three kinds of work

Pick the family by the kind of work. Then pick the tier by how hard the task is.

1. Chat: knowledge and documents. For questions, writing, research and analysis. Chat reads images, documents and PDFs as well as text. It has a 256K-token context window.
2. Coding: math and software. Optimized for math and code. Its 1M-token context window is the largest of the three families, so a whole repository fits in one request. That matters most on the hardest tasks.
3. Agents: volume work, 24/7. For agents that run all day and all night. Optimized for speed and price. It has a 256K-token context window.

Context is how much a model can read in one request. 1,000 tokens is about 750 words, so 256K tokens is about 190,000 words and 1M tokens is about 750,000.

### Fast, pro or ultra

Every family has the same three tiers. The tier sets how much intelligence you pay for on a task.

Figure. Example at list prices: one agent task of three calls, each with 4,000 tokens in and 500 tokens out, costs $0.0051 on agents fast, $0.0211 on agents pro and $0.0464 on agents ultra.
Alt text: Three steps rising from left to right. Fast: speed and low cost, for routine work in volume. Pro: balanced, for everyday work of medium difficulty. Ultra: the hardest tasks, where intelligence matters most.
- fast: Speed and low cost. Choose it when: The work is routine and comes in volume: rewriting, classifying, extracting, inner-loop tool calls.
- pro: Balanced. Choose it when: The work is everyday production of medium difficulty: summaries, code review, standard agents.
- ultra: The hardest tasks, where intelligence matters most. Choose it when: The task is deep research, system design or a long multi-step plan.

Most work does not need ultra. Some work does. If you are not sure, start with pro and compare it with fast on your own tasks.

- Estimate your own task (https://avelin.ai/pricing#estimate)

### All nine models

One table, every model. Prices are in US dollars per 1M tokens, at list price, as of 3 Oct 2026.

The nine AVELIN models by family and tier, with context window, list prices per 1M tokens and best use.
| Family | Tier | Model | Context | Input | Output | Best for |
| --- | --- | --- | --- | --- | --- | --- |
| Chat | fast | avelin-fast | 256K | $0.30 | $1.20 | High-volume drafting, classification and extraction |
| Chat | pro | avelin-pro | 256K | $1.40 | $4.40 | Everyday production: summaries, planning, internal Q&A |
| Chat | ultra | avelin-ultra | 256K | $2.50 | $12.50 | Deep research, analysis and high-stakes decisions |
| Coding | fast | avelin-coding-fast | 1M | $0.30 | $1.20 | Day-to-day coding and debugging |
| Coding | pro | avelin-coding-pro | 1M | $1.40 | $4.40 | Code review, multi-file changes and CI pipelines |
| Coding | ultra | avelin-coding-ultra | 1M | $2.50 | $12.50 | System design, security review and large refactors |
| Agents | fast | avelin-agentic-fast | 256K | $0.285 | $1.14 | High-volume agents and bulk classification |
| Agents | pro | avelin-agentic-pro | 256K | $1.26 | $3.96 | Standard tool-calling agents |
| Agents | ultra | avelin-agentic-ultra | 256K | $2.38 | $11.88 | Complex multi-step planning and long-running agents |

Every model supports tool calling, streaming and prompt caching, and can write a maximum of 65K tokens in one answer. Chat models accept images. Coding and Agents models take text. The Agents family uses the id prefix avelin-agentic. Older names such as avelin-coding and avelin-agentic-high still work as aliases.

- Cache rates and the full reference (https://avelin.ai/pricing)
- Model catalog (https://avelin.ai/docs/models)
- The same prices as JSON (https://api.avelin.ai/public/models.json)

Prefer a chat window to an API? The AVELIN App runs on the same nine models.

- AVELIN App (https://avelin.ai/private-gpt)

### How the nine models are built

Each of the nine models is fine-tuned for one kind of work and tuned to work with the Decision Layer. Our research direction for the next AVELIN models is cross-model fusion: combining the strengths of several models into one new model. It is research, not a shipped model.

Fusion belongs to model creation, before any request arrives. At runtime, the Decision Layer picks one model for each request, and that model answers. Nothing is fused when you send a request.

### Run them where you need

- AVELIN Cloud Call the models through one API and pay per token at list prices.
  - See the API platform (https://avelin.ai/api-platform)
- Your own GPUs AVELIN inference is optimized at GPU level for these nine models, and the platform installs on your hardware.
  - AVELIN on your GPUs (https://avelin.ai/sovereign-ai-platform)

- Not sure which fits? Compare the three deployments (https://avelin.ai/deployment)

### Tune them to your work

For enterprise customers, AVELIN can fine-tune or train the models on your own workflows, so they fit your documents, formats and tasks. We scope the work with you in a proof of concept (POC).

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

### Built to work with the Decision Layer

Teams send every request to the top model because nobody wants to risk a cheaper one. The Decision Layer makes that choice for each request. It reads the task, then picks the family and tier that fit: ultra when the task needs it, fast when it does not.

The nine models are tuned to work with the Decision Layer. They also work without it. You can name a tier yourself at any time.

The Decision Layer works with your own providers too, so a POC can start with the models you use today.

- See how the Decision Layer works (https://avelin.ai/decision-layer)

### FAQ: Questions about the models

**What is cross-model fusion?**
It is AVELIN's research direction for building new models: combining the strengths of several models into one. It is model creation, not request handling. At runtime the Decision Layer picks one model on every request, and that model answers.

**What are the AVELIN model families?**
There are three families. Chat is for knowledge work and documents, Coding is for math and software, and Agents is for volume work that runs 24/7. Each model is specially fine-tuned for its task type. Each family has three tiers, so there are nine models in all.

**Should I choose fast, pro or ultra?**
Choose fast for speed and low cost, pro for balanced work of medium difficulty, and ultra for the hardest tasks where intelligence matters most. If you are not sure, start with pro and compare it with fast on your own tasks.

**Which AVELIN model is best for AI agents?**
Use the Agents family. It is optimized for speed and price, so it suits volume work that runs all day and night. It can also do complex tasks, including coding, with a smaller context window than the Coding family.

**How much do the AVELIN models cost?**
You pay per token at public list prices, with no subscription and no minimum. Input prices run from $0.285 to $2.50 per 1M tokens, and output prices from $1.14 to $12.50. The pricing page lists every model with its cache rates.
More: Pricing (https://avelin.ai/pricing)

**How big is the context window?**
Chat and Agents models read 256K tokens in one request. Coding models read 1M tokens, enough for a whole repository. As a guide, 1,000 tokens is about 750 words.

**Can I tune the models to my workflows?**
Yes. For enterprise customers, AVELIN can fine-tune or train the models on your own workflows, and we scope it in a proof of concept (POC).

**Can I use the models without the Decision Layer?**
Yes. Name a tier in the model field of the API, and you pay that tier's list price. The Decision Layer is optional: it picks a tier for every request, so you pay for top intelligence only when a task needs it.
More: Decision Layer (https://avelin.ai/decision-layer)

**Can the models run on my own GPUs?**
Yes. The AVELIN platform installs on your GPUs, and its inference layer is optimized at GPU level for these nine models.
More: AVELIN on your GPUs (https://avelin.ai/sovereign-ai-platform)

Related: Decision Layer (https://avelin.ai/decision-layer); API platform (https://avelin.ai/api-platform); Sovereign AI platform (https://avelin.ai/sovereign-ai-platform); Pricing (https://avelin.ai/pricing); Deployment (https://avelin.ai/deployment)

### Find the right tier for your work.

A POC runs your workload through the Decision Layer and reports the cost per task next to your current bill.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## API platform
URL: https://avelin.ai/api-platform
Markdown copy: https://avelin.ai/api-platform.md
Updated: 3 Oct 2026

**One API. OpenAI and Anthropic compatible.**

Point your OpenAI or Anthropic client at AVELIN. Change the base URL and the key. You get nine models, speech, web search and web scraping on one key, with a cost view for every key.

- Start building (https://avelin.ai/docs/api/quickstart)
- Read the API reference (https://avelin.ai/docs/api)

Base URL: https://api.avelin.ai

### Change two lines

AVELIN speaks the OpenAI and Anthropic wire formats, so most existing SDKs and tools keep working. Change the base URL and the API key. Then name a tier, such as avelin-pro, in the model field. To switch back, change them back.

OpenAI SDK
```python
from openai import OpenAI

client = OpenAI(
    base_url="https://api.avelin.ai/v1",  # new base URL
    api_key="sk-avelin-YOUR_KEY",         # new API key
)

resp = client.chat.completions.create(
    model="avelin-pro",
    messages=[{"role": "user", "content": "..."}],
)
```

Anthropic SDK
```python
from anthropic import Anthropic

client = Anthropic(
    base_url="https://api.avelin.ai",  # new base URL
    api_key="sk-avelin-YOUR_KEY",      # new API key
)

msg = client.messages.create(
    model="avelin-pro",
    max_tokens=1024,
    messages=[{"role": "user", "content": "..."}],
)
```

curl, OpenAI format
```bash
curl https://api.avelin.ai/v1/chat/completions \
  -H "Authorization: Bearer $AVELIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "avelin-pro",
    "messages": [
      { "role": "user", "content": "Summarize the benefits of vendor-neutral AI." }
    ]
  }'
```

curl, Anthropic format
```bash
curl https://api.avelin.ai/v1/messages \
  -H "Authorization: Bearer $AVELIN_API_KEY" \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "avelin-coding-fast",
    "max_tokens": 1024,
    "messages": [
      { "role": "user", "content": "Write a Python function to debounce calls." }
    ]
  }'
```

Tool calling
```json
{
  "model": "avelin-agentic-pro",
  "messages": [
    { "role": "user", "content": "What's the weather in Abu Dhabi?" }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get current weather for a city",
        "parameters": {
          "type": "object",
          "properties": { "city": { "type": "string" } },
          "required": ["city"]
        }
      }
    }
  ],
  "tool_choice": "auto"
}
```
Streaming, tool calling and thinking control work on both formats. On the Anthropic format, send "thinking": { "type": "disabled" } to turn thinking off. On the OpenAI format, send "reasoning_effort": "high" to ask for deeper reasoning. Keys in the samples are placeholders.

- SDKs and tools (https://avelin.ai/docs/api/sdks)

### Two chat formats, one set of models

Use the format your code already speaks. Both reach the same nine models and return equivalent results.

Chat, model, embedding and image endpoints of the AVELIN API, with the model and list price of each.
| Endpoint | What it does | Model | List price |
| --- | --- | --- | --- |
| POST /v1/chat/completions | Chat in the OpenAI format, with streaming, tool calling and thinking control. | The nine AVELIN models | Per token |
| POST /v1/messages | Chat in the Anthropic format, with streaming, tool calling and thinking control. | The nine AVELIN models | Per token |
| GET /v1/models | Lists the models your key can use. | n/a | n/a |
| POST /v1/embeddings | Vector embeddings for search and retrieval. | bge-m3 | $0.02 per 1M tokens |
| POST /v1/images/generations | Image generation. | avelin-imagegen, avelin-imagegen-pro | $0.03 or $0.075 per image |

Model prices per token are on the price list. The nine models are described on the models page.

- Price list (https://avelin.ai/pricing)
- The nine models (https://avelin.ai/models)

### Agentic endpoints

Agents listen, speak and read the web. These endpoints use the same key and the same bill as your model calls.

Speech, web search and web scraping endpoints of the AVELIN API, with what each returns, the models and the list price.
| Endpoint | What you get | Models | List price |
| --- | --- | --- | --- |
| POST /v1/audio/transcriptions | Speech to text. Send audio, get text. | whisper-large-v3 for accuracy, whisper-large-v3-turbo for volume, avelin-stt for streaming-friendly work | $0.006 per minute |
| POST /v1/audio/speech | Text to speech. Turn text into audio. OpenAI voice names and three output formats. | tts-1 for low latency, tts-1-hd for the highest quality | $0.12 or $0.24 per 1K characters |
| POST /v2/search | Web search. Ranked results, with the result pages scraped in the same call if you ask. | n/a | $0.005 per request |
| POST /v2/scrape | Web scraping. One URL in, clean text out, ready for a model. | n/a | $0.001 per request |
| POST /v2/map, /v2/crawl, /v2/extract | Whole sites. Map a site's links, crawl it, or extract structured data. Crawl and extract run as jobs and are charged once, at the start. | n/a | $0.001, $0.010 and $0.005 per request |

List prices in US dollars, as of 3 Oct 2026. The web endpoints follow the Firecrawl v2 format and work with its SDK.

- Web endpoints reference (https://avelin.ai/docs/api/web-scraping)
- Full price reference (https://avelin.ai/docs/pricing)

### See what every key costs

Every request is counted against the key that sent it. Give each team, app or environment its own key, and the console shows usage and spend for each one, by model, as it happens.

1. Usage by key and model. Live in the console. Each record shows the key, the model, token counts, speed and status.
2. A budget on every key. Set a cap for each team, app or environment. Budget alerts warn you as a key fills.
3. Limits that act first. A key's budget and rate limits are applied before a request is answered. A key at its limit gets a 429 error.

Figure. Cost per task is counted as on the Method page: calls per task times the price of each call.
Alt text: Schematic of the cost view in the AVELIN console. Spend is split by key and by model tier, and each key has a dashed budget line. Example only, not real data.

- Method (https://avelin.ai/method)

### User management and scoped keys

Your team signs in to the web console at platform.avelin.ai. Create a separate key for each team, app or environment. Each key has its own budget and rate limits, and each can be rotated on its own.

Treat keys as secrets. Keep them out of client-side code and out of source control.

- Create a key in the console (https://platform.avelin.ai)

### How a request flows

Figure. Step 3 is yours when you name a tier, and the Decision Layer's when you use it.
Alt text: A row of five steps: your app sends a request, the key is checked, a model is chosen, the model answers, and the request is recorded.
1. Your app sends a request. It is the same call you make today, sent to api.avelin.ai with your key.
2. The key is checked. The key's budget and rate limits are applied before a request is answered.
3. A model is chosen. You name a tier, or the Decision Layer picks one for the request. Hard tasks go to a more capable model and easy tasks to a cheaper one.
4. The model answers. The answer comes back in the format you called, OpenAI or Anthropic, and streams if you ask. If a model hits a temporary error, AVELIN retries and can fail over to a backup. The price does not change.
5. The request is recorded. The usage record holds the key, the model, token counts, speed and status. Billing follows the list price of the tier.

### FAQ: Questions about the API

**Is the AVELIN API OpenAI compatible?**
Yes. Send chat requests to /v1/chat/completions with the OpenAI SDK or any tool that speaks the OpenAI format. Streaming and tool calling work, and embeddings, speech and image endpoints use the usual OpenAI paths.

**Is the AVELIN API Anthropic compatible?**
Yes. Send requests to /v1/messages with the Anthropic SDK. Streaming and tool calling work, and you control thinking on each request, for example by turning it off.

**How do I switch from my current provider?**
Change two things: the base URL and the API key. Then put an AVELIN tier, such as avelin-pro, in the model field. To switch back, change them back.
More: Quickstart (https://avelin.ai/docs/api/quickstart)

**Which agentic endpoints are included?**
Speech to text, text to speech, web search and web scraping, on the same key and the same bill as your model calls. Web scraping also covers map, crawl and extract, and the web endpoints follow the Firecrawl v2 format.
More: Web endpoints reference (https://avelin.ai/docs/api/web-scraping)

**How do I cap spend?**
Create a key for each team, app or environment, and set a budget and rate limits on it. The console shows usage and spend by key and by model, and a key that reaches its limit gets a 429 error.

**What does the API cost?**
Model calls are billed per token at public list prices, with no subscription and no minimum. Speech is priced per minute or per 1K characters, images per image, and web endpoints per request. The price reference lists every endpoint.
More: Price reference (https://avelin.ai/docs/pricing); Pricing (https://avelin.ai/pricing)

**Do you keep my prompts?**
AVELIN Cloud has zero retention by default, so we do not store your prompts or answers. We keep only the usage records needed to bill you and keep the service secure. Your prompts are not used to improve AVELIN unless you opt in. The data policy has the detail.
More: Data policy (https://avelin.ai/legal/data-policy); Privacy (https://avelin.ai/privacy)

**Can I use the Decision Layer with this API?**
Yes. The Decision Layer picks the model for every request, so you pay for top intelligence only when a task needs it. It works with AVELIN models and with your own providers, and you start with a proof of concept (POC).
More: Decision Layer (https://avelin.ai/decision-layer)

Related: Models (https://avelin.ai/models); Decision Layer (https://avelin.ai/decision-layer); Pricing (https://avelin.ai/pricing); Docs (https://avelin.ai/docs); Sovereign AI platform (https://avelin.ai/sovereign-ai-platform)

### Make your first request.

Or start a POC and see the cost per task on your own workload.

- Start building (https://avelin.ai/docs/api/quickstart)

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

## Deployment
URL: https://avelin.ai/deployment
Markdown copy: https://avelin.ai/deployment.md
Updated: 3 Oct 2026

**Cloud, enterprise or hybrid.**

Run AVELIN in our cloud, on your own GPUs, or both. You choose where the models run and what can leave. The Decision Layer picks the right model for every request in all three.

- Request a POC (https://avelin.ai/poc#request)
- Try the live demo (https://arena.avelin.ai)

### Pick where AVELIN runs

Three deployments. Choose by where your data must be processed.

- AVELIN Cloud. Best for teams that want to start now and cut token spend.
  - Nine AVELIN models and the Decision Layer behind one API
  - Connect your own AI providers
  - Zero retention by default
  - Start building (https://avelin.ai/docs/api/quickstart)
  - Read about the cloud (https://avelin.ai/deployment#cloud)
- Enterprise. Best for teams whose data must be processed on their own GPUs.
  - The Decision Layer, the full platform, or both
  - Inference optimized for your GPUs
  - Models tuned to your workflows
  - Request a POC (https://avelin.ai/poc?topic=enterprise#request)
  - Read about enterprise (https://avelin.ai/deployment#enterprise)
- Hybrid. Best for teams that want local control and extra intelligence on demand.
  - Local models on your GPUs
  - Offload to a public model when needed
  - A privacy check applies your rule
  - Request a POC (https://avelin.ai/poc?topic=hybrid#request)
  - Read about hybrid (https://avelin.ai/deployment#hybrid)

### AVELIN Cloud

Use the AVELIN models and the Decision Layer through one API. Switch by changing the base URL and the API key.

1. The Decision Layer. It reads each request and picks the model it needs. Hard tasks go to a capable model. Easy tasks go to a cheaper one.
2. Your own providers. Connect the AI providers you already use. The Decision Layer helps you cut token spend across them.
3. Global inference. Requests are served from the United States, Europe and MENA. Isolated in-country locations are planned.
4. Zero retention by default. We do not keep your prompts or answers, and they are not used to improve AVELIN unless you opt in. The data policy has the detail.

- Start building (https://avelin.ai/docs/api/quickstart)

- See the three privacy levels (https://avelin.ai/privacy#levels)
- Read the data policy (https://avelin.ai/legal/data-policy)

### Enterprise

AVELIN on GPUs you control. Use the Decision Layer on your workflows, the full platform, or both.

1. The Decision Layer on your workflows. Put it in front of your apps and agents. It picks the right model for every request, from the models you allow.
2. The full platform on your GPUs. The AVELIN inference layer, the nine models and the API platform, installed on your hardware and optimized for your GPUs.
3. Tuned to your workflows. Models can be fine-tuned or trained on your own workflows, so they fit your tasks.
4. Your data, your network. Prompts are processed on your GPUs, which is the ultimate privacy. The written scope lists every network path the software opens.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

- What a scope covers (https://avelin.ai/private-deployment)
- The sovereign AI platform (https://avelin.ai/sovereign-ai-platform)

### Hybrid

Local models run on your GPUs. When a task needs more intelligence, or your local model is busy, AVELIN can send it to a public model. A privacy check lets you decide whether detected private data may leave.

#### The privacy check, step by step

1. Read. The Decision Layer reads each request and picks a model. Your local model is the default.
2. Decide. If the task needs more intelligence, or your local model is under heavy load, a public model becomes an option.
3. Check. Before anything leaves your GPUs, the privacy check looks for private data in the request.
4. Apply your rule. You decide. Detected private data may go to a public model, or it must stay on your GPUs.
5. Send. The request stays on your GPUs, or goes to a public model. Every offloaded request passes the check first.

Figure. Hybrid. A request reaches the Decision Layer on your side. The privacy check applies your rule, and the request takes one of two exits.
Alt text: Diagram of a hybrid deployment. A request enters the Decision Layer inside a box labeled "your GPUs", beside the local models. When a task needs more intelligence or the local model is busy, the request passes through a privacy check. The check applies your rule and sends the request down one of two exits. The exit "stays on your GPUs" returns it to the local models. The exit "goes to a public model" leaves the box.

- Request a POC (https://avelin.ai/poc?topic=hybrid#request)

- Privacy at each level (https://avelin.ai/privacy#levels)

### Where models run and where data goes

The same questions, answered three ways.

Comparison of AVELIN Cloud, enterprise and hybrid deployments. On a phone, each row becomes a labeled list.
| Item | AVELIN Cloud | Enterprise | Hybrid |
| --- | --- | --- | --- |
| Where the models run | AVELIN infrastructure | Your GPUs | Local models on your GPUs. Public models when you allow it. |
| Where requests are processed | The United States, Europe and MENA | On your GPUs, in your network | On your GPUs by default. At a public model after an offload. |
| What AVELIN keeps | Zero retention by default | Set in the written scope | Set in the written scope |
| Can a request leave? | It is served by AVELIN and the providers you connect | You decide. The scope lists every network path. | Only through the privacy check, by your rule |

### In-country locations

We are planning isolated locations inside a single country or region, so your data does not cross a border.

If your data must stay in one country or region, say so in your POC request. The scope states where your data is processed before any data is sent.

- Request a POC (https://avelin.ai/poc#request)

### FAQ: Questions about deployment

**Which deployment fits us?**
Choose AVELIN Cloud to start now and cut token spend. Choose enterprise if your data must be processed on your own GPUs. Choose hybrid if you want local models, extra intelligence from public models when needed, and your own rule for private data.

**What is hybrid?**
Hybrid runs local models on your GPUs. It can offload a request to a public model when a task needs more intelligence or your local model is under heavy load. A privacy check applies your rule first.

**What does the privacy check do?**
It looks for private data in a request before the request can leave your GPUs. You decide what happens: detected private data may go to a public model, or it must stay on your GPUs.

**When does a request go to a public model?**
When the task needs more intelligence than your local model offers, or when your local model is under heavy load. It goes only if the privacy check allows it under your rule.

**Where does cloud inference run?**
Requests are served from the United States, Europe and MENA. The data policy lists the exact countries. If your data must stay in one country or region, say so in your POC request.
More: Data policy (https://avelin.ai/legal/data-policy#where-inference-runs)

**What are isolated in-country locations?**
They are locations in a single country or region where your data does not cross a border. They are planned for AVELIN Cloud.

**What is a sovereign AI platform deployment?**
It is a deployment where you, not a vendor or a cloud, decide where the system runs and who can reach it. An AVELIN enterprise deployment is one.
More: Sovereign AI platform (https://avelin.ai/sovereign-ai-platform)

**How do we start?**
Request a POC on one of your workloads. The scope states where your data is processed before any data is sent.
More: Request a POC (https://avelin.ai/poc#request)

Related: Privacy (https://avelin.ai/privacy); Sovereign AI platform (https://avelin.ai/sovereign-ai-platform); Decision Layer (https://avelin.ai/decision-layer); Models (https://avelin.ai/models); API platform (https://avelin.ai/api-platform); Private deployment scope (https://avelin.ai/private-deployment); Pricing (https://avelin.ai/pricing)

### Choose where it runs. Prove it first.

A POC is a proof of concept on your own workload. Tell us which deployment you are weighing and we scope it with you.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Privacy
URL: https://avelin.ai/privacy
Markdown copy: https://avelin.ai/privacy.md
Updated: 3 Oct 2026

**Privacy is three questions.**

What do you keep? Where does it run? Can it leave? AVELIN answers each one at three levels: cloud, hybrid and on-premise. You choose the level for each workload.

- Request a POC (https://avelin.ai/poc#request)
- Read the data policy (https://avelin.ai/legal/data-policy)

### Why privacy matters more now

In a 2026 survey of 750 IT, security and AI leaders, 76% said they were highly or extremely concerned about data security and privacy. It ranked first among the AI concerns in the survey.

Source: AvePoint, State of AI 2026, research by Osterman. AvePoint sells data-governance software and commissioned the survey.

Agents make many requests for one task, so more of your data passes through a model. The question is not only who can read a prompt. It is where the prompt runs, what is kept, and whether it can leave. Banning AI does not remove the risk. Control does.

### Three levels of control

Choose the level for each workload. The higher the level, the more you decide.

Figure. Three levels of control. Each step up gives you more say over where a request runs and what can leave.
Alt text: A ladder with three steps, rising from bottom to top. Level 1, AVELIN Cloud: zero retention by default. Level 2, Hybrid: you decide what may leave. Level 3, On-premise: prompts are processed on your own GPUs, the ultimate privacy.
- Level 1, AVELIN Cloud: Zero retention by default.
- Level 2, Hybrid: You set the rule for private data.
- Level 3, On-premise: Prompts are processed on your own GPUs.

Privacy at each level. On a phone, each row becomes a labeled list.
| Level | What do you keep? | Where does it run? | Can it leave? |
| --- | --- | --- | --- |
| 1. AVELIN Cloud | Zero retention by default. Not used to improve AVELIN. | AVELIN infrastructure, in the United States, Europe and MENA. | Requests are served by AVELIN. If you connect your own providers, their terms apply to what they receive. |
| 2. Hybrid | Set in your written scope. | On your GPUs by default. At a public model after an offload. | Only through the privacy check. You decide whether detected private data may go to a public model or must stay on your GPUs. |
| 3. On-premise | Set in your written scope. | On your GPUs, in your network. | You decide. The scope lists every network path the software opens. |

### AVELIN Cloud: zero retention by default

We do not keep the content of your prompts or answers in normal operation. Each request is processed in memory, and your prompts are not used to improve AVELIN.

We keep only the records needed to bill you and keep the service secure: timestamps, the model used, token counts, speed, status codes and your API key. None of it includes the content of your requests.

- Read the data policy (https://avelin.ai/legal/data-policy)
- App privacy policy (https://avelin.ai/legal/app-privacy)

### Opt in to help improve AVELIN

You can choose to share anonymized requests. They help us improve AVELIN, including how the Decision Layer picks a model.

What we keep is vector representations (embeddings) of your requests, with account identifiers removed. They are not linked to your account.

- Off by default. Nothing is shared unless you turn it on.
- Turn it off at any time. New sharing stops at once.
- The price terms are set in your agreement.

### Hybrid: you decide what may leave

In a hybrid deployment, local models run on your GPUs. When a request needs a public model, a privacy check looks for private data first. You decide what happens next: detected private data may go to a public model, or it must stay on your GPUs.

Figure. Hybrid. The privacy check applies your rule before a request can leave your GPUs.
Alt text: Small diagram of a hybrid deployment. Inside a box labeled "your GPUs", a privacy check sits between the Decision Layer and two exits: "stays on your GPUs" and "goes to a public model".

- How hybrid works, step by step (https://avelin.ai/deployment#hybrid)

### On-premise: the ultimate privacy

AVELIN runs on your GPUs, in your network. Prompts are processed on your hardware, and you decide where it runs and who can reach it.

This is what a sovereign AI platform means: you, not a vendor or a cloud, decide where the system runs and who can reach it. The written scope lists every network path the software opens.

- The sovereign AI platform (https://avelin.ai/sovereign-ai-platform)
- What a scope covers (https://avelin.ai/private-deployment)

### FAQ: Questions about privacy

**What is zero retention?**
Zero retention means AVELIN Cloud does not store the content of your prompts or answers in normal operation. Each request is processed in memory, and we keep only the records needed to bill you and keep the service secure. The data policy has the detail.
More: Data policy (https://avelin.ai/legal/data-policy)

**Do you train on my prompts?**
No. Your prompts and answers are not used to improve AVELIN unless you turn on sharing, and you can turn it off at any time.

**How do I switch sharing off?**
Turn it off in your account settings at platform.avelin.ai. New sharing stops at once.

**Where is my data processed?**
In AVELIN Cloud, requests are processed in the United States, Europe and MENA. Isolated in-country locations are planned. If your data must stay in one country or region, say so in your POC request. On-premise, requests are processed on your own GPUs.

**What does the hybrid privacy check do?**
Before a request leaves your GPUs, the privacy check looks for private data. You decide what happens: detected private data may go to a public model, or it must stay on your GPUs.
More: How hybrid works (https://avelin.ai/deployment#hybrid)

**Does zero retention cover the providers that serve my requests?**
AVELIN's infrastructure subprocessors are bound by data processing agreements, and a current list is available on request. If you connect your own providers to the Decision Layer, their terms apply to what they receive.
More: Subprocessors, in the data policy (https://avelin.ai/legal/data-policy#subprocessors)

Related: Deployment (https://avelin.ai/deployment); Sovereign AI platform (https://avelin.ai/sovereign-ai-platform); Decision Layer (https://avelin.ai/decision-layer); Trust (https://avelin.ai/trust); Data policy (https://avelin.ai/legal/data-policy); Pricing (https://avelin.ai/pricing)

### Bring your privacy requirements.

Tell us what you must keep, where it must run and what may leave, and we scope a POC around it.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Pricing
URL: https://avelin.ai/pricing
Markdown copy: https://avelin.ai/pricing.md
Updated: 3 Oct 2026

**Pricing.**

Nine models at public list prices. Pay per token. No subscription, no minimum.

### Price list

AVELIN list prices in US dollars per 1M tokens, by model
| Model | Family and tier | Context | In | Out | Cache write | Cache read |
| --- | --- | --- | --- | --- | --- | --- |
| avelin-fast | Chat, fast | 256K | $0.30 | $1.20 | $0.38 | $0.03 |
| avelin-pro | Chat, pro | 256K | $1.40 | $4.40 | $1.75 | $0.26 |
| avelin-ultra | Chat, ultra | 256K | $2.50 | $12.50 | $3.13 | $0.52 |
| avelin-coding-fast | Coding, fast | 1M | $0.30 | $1.20 | $0.38 | $0.03 |
| avelin-coding-pro | Coding, pro | 1M | $1.40 | $4.40 | $1.75 | $0.26 |
| avelin-coding-ultra | Coding, ultra | 1M | $2.50 | $12.50 | $3.13 | $0.52 |
| avelin-agentic-fast | Agents, fast | 256K | $0.285 | $1.14 | $0.36 | $0.0285 |
| avelin-agentic-pro | Agents, pro | 256K | $1.26 | $3.96 | $1.58 | $0.23 |
| avelin-agentic-ultra | Agents, ultra | 256K | $2.38 | $11.88 | $2.98 | $0.49 |

USD per 1M tokens, list price, as of 3 Oct 2026. A repeated prompt prefix is billed once at the cache write rate and then at the lower cache read rate. The API publishes the same prices as machine-readable JSON. Audio, image and web endpoints are in the full reference.

- machine-readable JSON (https://api.avelin.ai/public/models.json)
- full reference (https://avelin.ai/docs/pricing)

Chat and coding share list prices, and coding has the longer context window. The agents tiers have their own, slightly lower, list prices. The nine models are described on the models page.

- models page (https://avelin.ai/models)

### Decision Layer and enterprise

The Decision Layer is priced per request and quoted in a POC. Sharing is opt-in and off by default: how sharing works.

- how sharing works (https://avelin.ai/privacy#opt-in)

Enterprise and on-premise deployments are priced in a written scope: sovereign AI platform, private deployment.

- sovereign AI platform (https://avelin.ai/sovereign-ai-platform)
- private deployment (https://avelin.ai/private-deployment)

### Estimate a task

Cost per task is the formula on the Method page. Enter the shape of one task, and the table prices it at each list price.

- Method (https://avelin.ai/method)

- Estimate a task (https://avelin.ai/pricing#estimate)

### Price a log of your own tasks

Have a log of your own model calls? Load it and it is priced at each list price, per call and per task, beside what you paid if the log has a cost column. It is the same arithmetic as the estimator, run on your numbers, in your browser. A POC starts from a sample like this one.

- Price a log of your own tasks (https://avelin.ai/pricing#log)

### Controls

1. A budget and rate limits on every key. A workload or a team can be capped on its own key. A key's budget and rate limits are applied before a request is answered.
2. Live usage and spend. Usage is recorded per key, with budget alerts.
3. One list price per model. You pay the list price of the model that answers.

### Billing terms

These are the billing terms in the API terms of service, which control if this page differs.

- API terms of service (https://avelin.ai/legal/api-terms)

- Usage is billed per token at the published prices in effect at the time of use.
- Usage is charged at the end of each billing period or drawn from prepaid credits, depending on how your account is set up.
- Fees are exclusive of taxes.
- Fees and prepaid credits are non-refundable, except where an order is cancelled or the law requires a refund.
- Published prices and payment terms can change for future use, and a change takes effect when it is posted. A change does not affect usage that has already happened.
- There is no guaranteed service level unless one is agreed in a separately signed agreement.

For invoicing, purchase orders, volume terms or committed-use pricing, ask in your POC request: a signed agreement controls over these terms, and a POC scope states the prices it uses. Private deployments are priced per deployment: Private deployment.

- Private deployment (https://avelin.ai/private-deployment)

Model calls are billed at the list price of the tier that answers. The Decision Layer is priced separately, per request, and quoted in a POC.

### Next step

Send one workload and we price it at list beside your current bill. Or start on your own: change the base URL, and leave by changing it back.

- Request a POC (https://avelin.ai/poc#request)
- Get API access (https://avelin.ai/docs/api/quickstart)

### FAQ: Questions about pricing

**What do the AVELIN models cost?**
You pay per token at public list prices, with no subscription and no minimum. Input prices run from $0.285 to $2.50 per 1M tokens, and output prices from $1.14 to $12.50, as of 3 Oct 2026.
More: The nine models (https://avelin.ai/models)

**How is cost per task counted?**
Cost per task is the sum of every model call a task makes, each priced at list price: tokens in and out multiplied by the published rate, with cached tokens at the cache rate. Agents make many calls per task, so price per token alone hides the bill.
More: Method (https://avelin.ai/method)

**What are cache write and cache read rates?**
A repeated prompt prefix is billed once at the cache write rate and then at the lower cache read rate. Both rates are in the price list for every model.

**What does a deployment on our own GPUs cost?**
Enterprise and on-premise deployments are priced in a written scope. AVELIN Cloud has public list prices.
More: Sovereign AI platform (https://avelin.ai/sovereign-ai-platform); Private deployment (https://avelin.ai/private-deployment)

Related: Models (https://avelin.ai/models); API platform (https://avelin.ai/api-platform); Decision Layer (https://avelin.ai/decision-layer); Method (https://avelin.ai/method)

### Bring one workload. See its cost per task.

Start with a POC on one of your workloads.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Request a POC
URL: https://avelin.ai/poc
Markdown copy: https://avelin.ai/poc.md
Updated: 3 Oct 2026

**A POC on one workload.**

A POC is a proof of concept on one of your workloads. We run the Decision Layer on it and show the cost per task next to your current setup.

- Request a POC (https://avelin.ai/poc#request)

Runs beside your current setup. Nothing in your production path.

### How a POC runs

1. Your workload. You share a sample of your AI tasks: real tasks with the sensitive parts removed, real tasks under an NDA, or made-up tasks. You choose.
2. Our Decision Layer. It reads each task and picks a model from the providers you connect, next to the choice you make today.
3. Side by side. You get a cost-per-task report. It shows your current bill and the Decision Layer's, task by task, at list prices.

### What you send and what you get

#### What you send

- Token counts and bills only. No task text leaves your side. Use the log format of the log pricer on the price page.
- Made-up tasks, or real tasks with the sensitive parts removed. The Decision Layer reads the opening of each request, so its choice for each task needs the text.
- Real tasks, under an NDA. If your data location allows it.

#### What you get

A cost-per-task report on your own workload at list prices. It has one row per task: your current bill, what the same tokens would cost on each AVELIN tier, and the choice the Decision Layer made, with totals at the end. It also reports the latency the Decision Layer adds on your sample, and gives a price for the Decision Layer, per request.

#### How it is scoped

We reply by email with a one-page scope. It states the workload, the data we need and where it is processed, how long it is kept, which terms apply and how long the POC runs. If your data cannot leave your country or network, say so in the data location field. A POC for an enterprise or hybrid deployment starts from the same form.

- How we measure (https://avelin.ai/method)
- Privacy (https://avelin.ai/privacy)
- Price a task yourself first (https://avelin.ai/pricing#estimate)

### FAQ: Questions about a POC

**What is a POC?**
A POC is a proof of concept on one of your workloads. We run the Decision Layer on a sample of your tasks, beside your current setup, and show the cost per task side by side at list prices.

**What do you need from us?**
A sample of your AI tasks, and the bill you pay for them today. You choose how much to share: token counts and bills only, made-up tasks, real tasks with the sensitive parts removed, or real tasks under an NDA.

**Does a POC touch our production path?**
No. It runs beside your current setup, so nothing in your production path changes and nothing is switched over.

**What do we get?**
A cost-per-task report. It shows your current bill and the Decision Layer's, task by task, at list prices, with the tasks where the Decision Layer costs more listed first. It also reports the latency the Decision Layer adds on your sample, and gives a price for the Decision Layer, per request.

**Who sees our data?**
The written scope states what data we need, where it is processed and how long it is kept, before any data is sent. If you ask for an NDA, it is signed before you share anything about the workload.
More: Privacy (https://avelin.ai/privacy); Trust (https://avelin.ai/trust)

**What happens next?**
We reply by email with a one-page scope. Nothing starts, and nothing is charged, before you have read it and accepted it. If the numbers hold, you can move to AVELIN Cloud, or run on your own GPUs.

Related: Decision Layer (https://avelin.ai/decision-layer); Deployment (https://avelin.ai/deployment); Pricing (https://avelin.ai/pricing); How we measure (https://avelin.ai/method); Trust (https://avelin.ai/trust)

### Not ready to send a workload?

Look around first. The live demo runs a task on a flagship model and on AVELIN, and the API takes two lines to try.

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Company
URL: https://avelin.ai/company
Markdown copy: https://avelin.ai/company.md
Updated: 3 Oct 2026

**A deep-tech AI company.**

A team of scientists and engineers doing R&D, training models and building the product. We make the Decision Layer, which picks the right model for every request, and the sovereign AVELIN platform, which runs on your GPUs.

### Research and development

AVELIN is deep tech. Research, model training and product engineering happen in one team.

1. The decision model. We train the model at the heart of the Decision Layer. It reads a request and picks the model the task needs. We also build the data it learns from and the evaluation that checks it. How we measure (https://avelin.ai/method)
2. Inference. We build the inference layer that serves the nine AVELIN models, and we optimize it at GPU level, for our cloud and for customer hardware.
3. The nine models. 3 families · 9 models for chat, coding and agents, each at fast, pro and ultra, tuned to work with the Decision Layer. For enterprises, we fine-tune or train models on their workflows.
4. Measurement. We publish how cost per task is counted, the recorded runs and the live Arena, including the runs where AVELIN cost more.

### What we build

1. AVELIN Decision Layer. Picks the right model for every request. Works with the providers you already use. A POC tests it on your workload. Decision Layer (https://avelin.ai/decision-layer)
2. Sovereign AVELIN platform. An inference layer on your GPUs, nine AVELIN models and an API platform. AVELIN Cloud runs the same models if you want to start there. Sovereign AI platform (https://avelin.ai/sovereign-ai-platform)
3. Also from AVELIN. AVELIN App, a workspace for chat, code and agents. Y-Ray Data, search over your documents and the web, with sources. AVELIN App (https://avelin.ai/private-gpt); Y-Ray Data (https://avelin.ai/business/y-ray-data)

### What you can check without asking us

- The live Arena, where every run is listed, including the runs where AVELIN cost more: 122 recorded runs, 24 Sep to 2 Oct 2026. Try the live demo (https://arena.avelin.ai)
- The prices, as the API publishes them, the same as the price list. Price list (https://avelin.ai/pricing)
- The cost formula and the recorded runs. How we measure (https://avelin.ai/method)
- The API itself: create a key and send a request. Start building (https://avelin.ai/docs/api/quickstart)
- The terms we operate under: the API terms and the data policy. Data policy (https://avelin.ai/legal/data-policy)

### Team

The team is listed on the Company page.

### Programmes and recognition

Programmes AVELIN takes part in.

- NVIDIA Inception. Selected as a member.
- Red Hat Partner Program. Accepted as a software partner.
- DigitalOcean Hatch. Member of the programme.
- DigitalOcean Partner Pod. Advanced to the ISV track, as a go-to-market partner.
- ElevenLabs. Awarded a startup grant.
- Oraseya Capital Launchpad. Joined the Launchpad programme of the venture fund of the Dubai Integrated Economic Zones Authority.
- AltaLab. Joined the fall cohort of the acceleration programme from the venture firm AltaIR Capital.
- Ignyte. Joined the Ignyte ecosystem, curated by the Dubai International Financial Centre.
- Business in Dubai. Takes part in the programme run by the Dubai Chamber of Digital Economy.
- Recognition. Inc. Arabia, Leaders of Tomorrow Awards, Dubai 2026: recognized in the AI category.
- Events. Money20/20 Middle East 2026, Riyadh. Web Summit 2026, Lisbon.

- Announcements are in News (https://avelin.ai/news)

### Company

AVELIN AI. Addresses are on the Contact page. Company announcements are in News.

- Contact (https://avelin.ai/contact)
- POC form (https://avelin.ai/poc#request)
- News (https://avelin.ai/news)
- AVELIN on LinkedIn (https://www.linkedin.com/company/avelin-ai)

### FAQ: Questions about the company

**Who builds AVELIN?**
A team of scientists and engineers doing R&D, training models and building the product. The team, with roles and background, is listed on this page.

**What does R&D cover?**
Four areas: the model that selects the right model for each request, inference optimized at GPU level, the nine AVELIN models, and measurement of cost per task. For enterprise customers, it also covers tuning or training models on their workflows.

**Where are you based?**
AVELIN AI L.L.C-FZ in Dubai signs the agreements and handles billing. The company works from the United States and the UAE. Addresses are on the Contact page.
More: Contact (https://avelin.ai/contact)

**Which programmes back you?**
AVELIN takes part in NVIDIA Inception, the Red Hat Partner Program, DigitalOcean Hatch and the DigitalOcean Partner Pod, Oraseya Capital Launchpad, AltaLab, Ignyte and Business in Dubai, and received a startup grant from ElevenLabs.

**How do I reach you?**
Write to hello@avelin.ai or use the form on the Contact page. To test AVELIN on a workload, request a POC.
More: Request a POC (https://avelin.ai/poc#request)

Related: How we measure (https://avelin.ai/method); Partners (https://avelin.ai/partners); News (https://avelin.ai/news); What is AVELIN (https://avelin.ai/what-is-avelin)

### Test it on your own workload.

Start with a POC and see your cost per task next to your current bill.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Private deployment
URL: https://avelin.ai/private-deployment
Markdown copy: https://avelin.ai/private-deployment.md
Updated: 3 Oct 2026

**Private deployment.**

AVELIN on GPUs you control. You decide where the system runs and who can reach it.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

### What the word means

Some buyers call this sovereign AI. Here it means one thing: you, not a vendor or a cloud, decide where the system runs and who can reach it.

A private deployment is AVELIN running on hardware you control. AVELIN Cloud is the same models and the Decision Layer, run by AVELIN, if you want to start there.

- Sovereign AI platform (https://avelin.ai/sovereign-ai-platform)
- Deployment (https://avelin.ai/deployment)

### What runs there

The AVELIN inference layer, optimized for your GPUs, the nine AVELIN models, the API platform and, if you want it, the Decision Layer. Which models your site runs, and on which GPUs, is settled in the written scope.

### What a scope settles

Every private deployment is scoped in writing before anything is installed. Each row is something the written scope states.

- Where it runs: The site and the hardware, on GPUs you control, in a network you control.
- What leaves: Every network path the software opens, such as updates or support access, and for each one whether it is on or off.
- Who has access: Who can reach the system, how any vendor support access is granted, and who approves it.
- Hardware: The GPU class and how many, for each tier you need.
- Updates: How new versions and models reach you, and who installs them.
- Licence: The terms on which you run the software and each model.
- Support: Who supports it, and when.
- Records: What is logged, where it is kept, and what we may see.
- Time and price: How long from agreement to a first install, and how it is priced.

To start, request a POC and say in the notes that you want AVELIN on your own hardware. The scope states who does what, what counts as success, and whether there is a fee.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

- Privacy (https://avelin.ai/privacy)
- Trust (https://avelin.ai/trust)

### FAQ: Questions about private deployment

**What is a private deployment?**
A private deployment is AVELIN running on hardware you control. You, not a vendor or a cloud, decide where the system runs and who can reach it. Some buyers call this sovereign AI.
More: Sovereign AI platform (https://avelin.ai/sovereign-ai-platform)

**What does a written scope settle?**
Where it runs, what leaves, who has access, hardware, updates, licence, support, records, and time and price. It lists every network path the software opens and whether each is on or off.

**What runs on our GPUs?**
The AVELIN inference layer, optimized for your GPUs, the nine AVELIN models, the API platform and, if you want it, the Decision Layer. Which models your site runs, and on which GPUs, is settled in the written scope.
More: Models (https://avelin.ai/models); API platform (https://avelin.ai/api-platform)

**How do we start?**
With a POC on one workload, then a written scope, then the install on your GPUs.
More: Request a POC (https://avelin.ai/poc?topic=enterprise#request)

**What does it cost?**
A deployment on your GPUs is priced in the written scope. AVELIN Cloud has public list prices.
More: Pricing (https://avelin.ai/pricing)

Related: Sovereign AI platform (https://avelin.ai/sovereign-ai-platform); Deployment (https://avelin.ai/deployment); Privacy (https://avelin.ai/privacy); Trust (https://avelin.ai/trust)

### Start with a POC.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Method
URL: https://avelin.ai/method
Markdown copy: https://avelin.ai/method.md
Updated: 3 Oct 2026

**How we measure.**

One task, two bills, one formula, printed. How AVELIN counts cost per task, what the live Arena shows, and how you can check it.

### The live Arena

In our live Arena, at list prices.

The Arena runs the same task on a flagship model and on AVELIN, then prices both. Over 122 recorded runs, 24 Sep to 2 Oct 2026, the median run cost 90.1% less on AVELIN. The 75th percentile was 94.4% and the best run 99.4%. Added up, the 122 runs cost 76.6% less.

It counts cost only. You read both answers yourself, so quality is judged separately. In 8 of the 122 runs AVELIN cost more. The comparison model differs by run, and every run is listed.

Median lower cost per task in the Arena, by the AVELIN model that answered
| AVELIN model | Runs | Median lower cost per task |
| --- | --- | --- |
| Chat fast | 43 | 94.1% |
| Chat pro | 54 | 92.5% |
| Chat ultra | 17 | 62.8% |
| Coding pro | 6 | 33.8% |
| Agents | 2 | Too few runs to report |

The Arena is a public demo with its own privacy page, so do not paste confidential data into it.

- All 122 runs, as a CSV (https://avelin.ai/arena-runs-2026-10-03.csv)
- The Arena scoreboard (https://arena.avelin.ai/scoreboard)
- Try the live demo (https://arena.avelin.ai)

### Cost per task

Cost per task is the sum of every model call a task makes, each priced at list price. Tokens in and tokens out are multiplied by the published rate for the model that answered. Cached tokens are priced at the cache rate, and reasoning tokens are billed as output. AVELIN list prices are on the price list, as of 3 Oct 2026. We use no negotiated prices, volume discounts or credits, on either side.

cost = (uncached input tokens × input price + cache read tokens × cache read price + cache write tokens × cache write price + output tokens × output price) ÷ 1,000,000

- Run the formula on your own numbers (https://avelin.ai/pricing#estimate)
- The price list (https://avelin.ai/pricing)

### The Decision Layer

AVELIN's own trained decision model reads each request and picks the model it needs: a more capable model when the task needs intelligence, a cheaper one when it is easy, on every request.

- Reads. The opening of a request, about the first 1,500 characters.
- Predicts. The kind of job, from 15 categories, and how hard it is, on a 5-level scale.
- Chooses. The model that fits, a more capable one for hard tasks and a cheaper one for easy tasks. A task of several steps can change model between steps.
- Records. Every decision is logged with its reason, so a choice can be checked afterwards against what each model would have cost.

- The Decision Layer (https://avelin.ai/decision-layer)
- Privacy (https://avelin.ai/privacy)

### Check it yourself

Price the AVELIN side of any task with the formula above. Send a prompt to the API, read the token counts in the response and multiply them by the list prices. Run the same prompt on the model you use today, then compare the two answers and the two bills.

curl
```bash
curl https://api.avelin.ai/v1/chat/completions \
  -H "Authorization: Bearer $AVELIN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "avelin-pro", "messages": [{"role": "user", "content": "Your prompt here"}]}'
```

In a POC the other side is your own bill, and the report is your own workload at list prices.

- Request a POC (https://avelin.ai/poc#request)

### FAQ: Questions about the method

**How is cost per task counted?**
Cost per task is the sum of every model call a task makes, each priced at list price. Tokens in and out are multiplied by the published rate for the model that answered, cached tokens are priced at the cache rate, and reasoning tokens are billed as output.

**What does the live Arena measure?**
In our live Arena, at list prices. The Arena runs the same task on a flagship model and on AVELIN and prices both. Over 122 recorded runs from 24 Sep to 2 Oct 2026, the median run cost 90.1% less. It counts cost only, and quality is judged separately: you read both answers yourself.

**What does the decision model read and predict?**
It reads the opening of a request, about the first 1,500 characters. It predicts the kind of job, from 15 categories, and how hard it is, on a 5-level scale. The Decision Layer then picks the model that fits and logs the reason.

**Can I check the figures myself?**
Yes. Every Arena run is listed in a public CSV and on the Arena scoreboard. You can also send a prompt to the API, read the token counts in the response and apply the formula on this page.

**How does a POC measure cost?**
A POC runs the Decision Layer on a sample of your own tasks and prices each task at list price next to your current bill. The report lists the tasks where the Decision Layer costs more first.
More: Request a POC (https://avelin.ai/poc#request)

Related: Decision Layer (https://avelin.ai/decision-layer); Pricing (https://avelin.ai/pricing); Trust (https://avelin.ai/trust); Company (https://avelin.ai/company)

### Measure it on your own tasks.

A POC runs the Decision Layer on a sample of your tasks and prices each one next to your current bill.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Trust
URL: https://avelin.ai/trust
Markdown copy: https://avelin.ai/trust.md
Updated: 3 Oct 2026

**Trust.**

What happens to your data, which controls exist and which terms apply. The plain-language version of the data side is on the Privacy page.

### The path of one request

1. Your application sends the request over TLS, with your key. Each key carries its own budget and rate limits.
2. AVELIN receives it in one of its processing regions. Cloud inference runs across the United States, Europe and MENA. Prompt and answer content is not stored in normal operation.
3. A model answers it. An AVELIN model, by task and tier, picked by you or by the Decision Layer.
4. The answer returns the same way. AVELIN keeps metadata to bill you and keep the service secure. Learning on your content is off unless you turn it on.

### Your data

1. Prompts and answers. Zero retention by default in AVELIN Cloud. They are not used to improve AVELIN unless you opt in, and you can withdraw consent at any time. Data policy (https://avelin.ai/legal/data-policy)
2. Records we keep. Metadata to bill you and keep the service secure: timestamps, the model used, token counts, speed, status codes and your API key. Security and network logs, which include IP addresses, are kept for a limited time that the data policy states. What we retain (https://avelin.ai/legal/data-policy#what-we-do-retain)
3. Where requests run. AVELIN Cloud serves requests across the United States, Europe and MENA. The data policy lists the exact countries. Customers with data-residency requirements can ask about pinning to one region. Where inference runs (https://avelin.ai/legal/data-policy#where-inference-runs)
4. Your own providers. If you connect your own providers to the Decision Layer, their terms apply to what they receive.
5. Subprocessors. Infrastructure subprocessors are bound by data processing agreements. A current list is available on request. Subprocessors (https://avelin.ai/legal/data-policy#subprocessors)

### Security and access

- Bearer-token API keys, with a budget and rate limits on each key.
- TLS on the API and on the web console.
- Usage records per key, with budget alerts.
- Encryption at rest for stored metadata, role-based access controls and network isolation between serving environments, as the API privacy policy states.

- Security and trust, in the docs (https://avelin.ai/docs/systems/avelin-security-trust)
- API privacy policy (https://avelin.ai/legal/api-privacy)

### Terms

1. API use. The API terms of service apply. Where you sign a separate agreement, such as a data processing agreement, it controls over the standard terms. The terms are governed by the laws of the United Arab Emirates. API terms (https://avelin.ai/legal/api-terms)
2. POC. A written scope for each engagement, with an NDA first if you ask for one. The scope says which terms apply. Nothing starts until you accept it. Request a POC (https://avelin.ai/poc#request)
3. Enterprise and private deployment. A data processing agreement, a service level, indemnity and governing law are set in the written scope. What a scope settles (https://avelin.ai/private-deployment)

### Questions

Send data-handling and security questions, or a security questionnaire, to hello@avelin.ai, or with your POC request. To report a security vulnerability, write to the same address with “Vulnerability” in the subject.

AVELIN AI. Addresses are on the Contact page.

- security.txt (https://avelin.ai/.well-known/security.txt)
- Contact (https://avelin.ai/contact)

- Request a POC (https://avelin.ai/poc#request)

### FAQ: Questions about trust

**Does AVELIN store my prompts?**
In AVELIN Cloud, zero retention is the default: prompts and answers are not stored in normal operation. The data policy lists the narrow exceptions and the periods that apply.
More: Data policy (https://avelin.ai/legal/data-policy); Privacy (https://avelin.ai/privacy)

**Does AVELIN train on my prompts?**
No, not by default. Your prompts are used to improve AVELIN only if you turn on sharing, and you can turn it off at any time.
More: How sharing works (https://avelin.ai/privacy#opt-in)

**Does AVELIN hold SOC 2, HIPAA or ISO certification?**
No. AVELIN does not hold SOC 2, HIPAA or ISO certification today and does not claim them. For a security review, use the data policy, the API terms and a security questionnaire, which you can send to hello@avelin.ai or with your POC request.

**Where are requests processed?**
AVELIN Cloud serves requests across the United States, Europe and MENA. The data policy lists the exact countries. If your data must stay in one country or region, say so in your POC request, and the scope names where it is processed before any data is sent.
More: Deployment (https://avelin.ai/deployment)

**Where do I send a security questionnaire?**
To hello@avelin.ai, or with your POC request.
More: Request a POC (https://avelin.ai/poc#request); Contact (https://avelin.ai/contact)

Related: Privacy (https://avelin.ai/privacy); Data policy (https://avelin.ai/legal/data-policy); Method (https://avelin.ai/method); Deployment (https://avelin.ai/deployment)

### Bring your review questions.

Send your security questionnaire with a POC request, and the written scope answers it before any data is sent.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Resources
URL: https://avelin.ai/resources
Markdown copy: https://avelin.ai/resources.md
Updated: 3 Oct 2026

**Resources**

Every page, doc and file on avelin.ai in one place. Start here to find a product, check a number, read a policy or download a logo.

### Products

- Decision Layer (https://avelin.ai/decision-layer): Picks the right model for every request.
- Sovereign AI platform (https://avelin.ai/sovereign-ai-platform): Inference, nine models and APIs on your GPUs.
- Models (https://avelin.ai/models): 3 families · 9 models, with context and prices.
- API platform (https://avelin.ai/api-platform): One API, OpenAI and Anthropic compatible.
- Deployment (https://avelin.ai/deployment): Cloud, enterprise or hybrid.
- Privacy (https://avelin.ai/privacy): Three questions, three levels of control.
- Pricing (https://avelin.ai/pricing): List prices and an estimator.
- AVELIN App (https://avelin.ai/private-gpt): A workspace for chat, code and agents.
- Y-Ray Data (https://avelin.ai/business/y-ray-data): Search over your documents and the web, with sources.
- Private deployment (https://avelin.ai/private-deployment): What a written scope settles.

### Start

- What is AVELIN? (https://avelin.ai/what-is-avelin): Plain answers on the whole company.
- Four problems enterprises hit with AI (https://avelin.ai/enterprise-ai-problems): Token price, the wrong model, budgets and privacy, explained with sources.
- Request a POC (https://avelin.ai/poc#request): Test the Decision Layer on your workload.
- Try the live demo (https://arena.avelin.ai): Run a task on a flagship model and on AVELIN, and compare the two bills.
- Start building (https://avelin.ai/docs/api/quickstart): Change the base URL and send a request.

### Business pages and use cases

#### Priced workloads (8)

#### Finance (10)

#### Healthcare (10)

#### Oil, Gas & Heavy Manufacturing (10)

#### Government, Defense & Public Sector (10)

#### Cross-Enterprise (10)

### Docs

- Documentation (https://avelin.ai/docs): Start here.
- API (https://avelin.ai/docs/api): Endpoints, SDKs and web data.
- API quickstart (https://avelin.ai/docs/api/quickstart): The two-line switch.
- Model catalog (https://avelin.ai/docs/models): Capabilities and a selection guide.
- Systems (https://avelin.ai/docs/systems): The parts of AVELIN and how a request flows.
- Benefits and advantages (https://avelin.ai/docs/benefits): How AVELIN creates measurable business value.
- Operations (https://avelin.ai/docs/operations): Running AVELIN reliably and securely.
- MCP (https://avelin.ai/docs/mcp): Actions in Gmail, Calendar, Contacts and Tasks.
- Integration guides: coding (https://avelin.ai/docs/guides/coding): Editors and coding agents.
- Integration guides: chat (https://avelin.ai/docs/guides/chat): Chat clients.
- Integration guides: automation (https://avelin.ai/docs/guides/automation): Agent and automation tools.
- Integration guides: creative (https://avelin.ai/docs/guides/creative): Creative tools.
- Integration guides: frameworks (https://avelin.ai/docs/guides/frameworks): Agent and app frameworks.
- Integration guides: observability (https://avelin.ai/docs/guides/observability): Tracing and monitoring tools.
- Platform overview (https://avelin.ai/docs/platform-overview): The platform in one document.
- Architecture (https://avelin.ai/docs/architecture): How the parts fit together.
- Glossary (https://avelin.ai/docs/glossary): Business and operating terms.
- Full price reference (https://avelin.ai/docs/pricing): Every model and endpoint at list prices.

#### Docs: Start (5)

#### Docs: API (5)

#### Docs: Benefits (6)

#### Docs: Integrations (66)

#### Docs: MCP (7)

#### Docs: Models (12)

#### Docs: Operations (4)

#### Docs: Systems (11)

#### Docs: Templates (1)

#### Docs: Use cases (5)

### How we know

- Method (https://avelin.ai/method): How cost per task is counted.
- Trust (https://avelin.ai/trust): What AVELIN states about data, and what it does not.
- The Arena (https://arena.avelin.ai): Every run is listed, including the runs where AVELIN cost more.
- Earlier sample, 60 runs of 26 Sep 2026, CSV (https://avelin.ai/recorded-runs-2026-09-26.csv)
- Earlier sample, prompts and answers (https://avelin.ai/recorded-runs-2026-09-26-prompts-and-answers.json)
- Machine-readable prices (https://api.avelin.ai/public/models.json)

### News and research

- News (https://avelin.ai/news): Company news, announcements and research essays.
- Company (https://avelin.ai/company): Who builds AVELIN, and what R&D covers.
- Contact (https://avelin.ai/contact): Write to AVELIN.

### Legal

- Data policy (https://avelin.ai/legal/data-policy)
- API privacy policy (https://avelin.ai/legal/api-privacy)
- API terms (https://avelin.ai/legal/api-terms)
- Website privacy policy (https://avelin.ai/legal/privacy)
- Website terms (https://avelin.ai/legal/terms)
- App privacy (https://avelin.ai/legal/app-privacy)
- App terms (https://avelin.ai/legal/app-terms)
- MCP privacy (https://avelin.ai/legal/mcp-privacy)
- MCP terms (https://avelin.ai/legal/mcp-terms)

### Brand assets

- Brand assets (https://avelin.ai/resources/brand): Logos, icons, favicons and colors.

### For machines

- llms.txt (https://avelin.ai/llms.txt): A short index for AI systems.
- llms-full.txt (https://avelin.ai/llms-full.txt): Full page text for AI systems.
- Sitemap (https://avelin.ai/sitemap.xml): Every indexable URL.
- Markdown copy of a page (https://avelin.ai/what-is-avelin.md): Add .md to the path of a page. This is the copy of What is AVELIN.

### FAQ: Questions about the hub

**What is the resources hub?**
A single page that links to every public page, doc and file on avelin.ai. Use it to find a product, check how a number was counted, read a policy or download brand assets.

**How can I check AVELIN's cost figures?**
Open the Method page to see how cost per task is counted. The live Arena lists every run, and the Method page has an earlier sample of 60 recorded runs, including the runs where AVELIN cost more.
More: Method (https://avelin.ai/method)

**Where are the API docs and prices?**
The docs start at the Documentation page. The API quickstart shows the two-line switch, and the full price reference lists every model and endpoint at list prices.
More: Documentation (https://avelin.ai/docs); Price reference (https://avelin.ai/docs/pricing)

**Where are the use cases?**
In the use-case index under Business. It lists the industry blueprints, and every use case, including eight priced workloads, is linked on this page.
More: Use cases (https://avelin.ai/business/use-cases)

**Where are the legal and privacy documents?**
All nine legal pages are in the Legal section above, including the data policy and the API terms.

**Where are the logos and brand colors?**
On the Brand assets page. It has logos, icons, favicons and the color palette.
More: Brand assets (https://avelin.ai/resources/brand)

### Test it on your own workload.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

## Brand assets
URL: https://avelin.ai/resources/brand
Markdown copy: https://avelin.ai/resources/brand.md
Updated: 3 Oct 2026

**Brand Resources**

The AVELIN brand kit: logos, icons, favicons, and brand colors. Grab the SVG or PNG you need, or copy a color straight from the palette.

The full page is at https://avelin.ai/resources/brand.

### FAQ: Questions about the brand kit

**Which logo files are available?**
Logos with text, the minimal circular mark and the letter A icon, each as SVG and PNG, in transparent, black, white and monochrome versions.

**Which colors does AVELIN use?**
Purple #B173E7, cyan #00EEFF, gold #F9C824 and a deep background, #0B0B0F. Each swatch on the page copies its HEX or RGB value.

**Which fonts does AVELIN use?**
Bitter for headings and IBM Plex Mono for interface text.

**Where do I get favicons and app icons?**
In the Favicons section: ICO, SVG, PNG, the Apple touch icon, the two web-app icons and the web manifest.

**Who do I ask about using the logo?**
Write to hello@avelin.ai.
More: Contact (https://avelin.ai/contact)

Related: Resources (https://avelin.ai/resources); Company (https://avelin.ai/company)

### Test it on your own workload.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

# Documentation

> The AVELIN docs, as plain text. Source: https://avelin.ai/docs

---

URL: https://avelin.ai/docs

# AVELIN Documentation

Welcome to AVELIN — a sovereign AI platform that gives you access to the world's best AI models through a single, governed API.

This documentation helps you get started, choose the right model, integrate with your tools, and understand the platform.

---

## 🚀 Quick Start

New to AVELIN? Start here.

- **[API Quickstart](api/quickstart.md)** — Make your first API call in under 5 minutes
- **[Platform Overview](platform-overview.md)** — What AVELIN is and how it works
- **[Pricing](pricing.md)** — Transparent per-token pricing for all models

---

## 🤖 AI Models

AVELIN offers 9 specialized models across 3 families, plus utility models. Each model is optimized for specific use cases.

### [Model Catalog](models/README.md)
Complete model comparison with context windows, max output, capabilities, and selection guidance.

#### Intelligence Models (256K context)
For reasoning, analysis, and multimodal tasks.

- **[avelin-ultra](models/avelin-ultra.md)** — Flagship reasoning for complex decisions and strategic analysis
- **[avelin-pro](models/avelin-pro.md)** — High-efficiency reasoning for daily workflows and production economics
- **[avelin-fast](models/avelin-fast.md)** — Low-latency responses for real-time chat and high-volume workloads

#### Coding Models (1M context)
For software engineering, debugging, and code generation.

- **[avelin-coding-fast](models/avelin-coding-fast.md)** — Balanced engineering tier for daily development
- **[avelin-coding-pro](models/avelin-coding-pro.md)** — High-throughput coding for production workloads
- **[avelin-coding-ultra](models/avelin-coding-ultra.md)** — Deep reasoning for system design and architecture

#### Agentic Models (256K context)
For autonomous agents, tool calling, and multi-step workflows.

- **[avelin-agentic-pro](models/avelin-agentic-pro.md)** — Standard agentic tier for tool-calling workflows
- **[avelin-agentic-ultra](models/avelin-agentic-ultra.md)** — Top-tier reasoning for complex multi-step planning
- **[avelin-agentic-fast](models/avelin-agentic-fast.md)** — Cost-optimized for high-volume agent swarms

#### Utility Models
Specialized models for embeddings, transcription, text-to-speech, and image generation.

- **[Utility Models](models/utility-models.md)** — Embeddings (bge-m3), transcription (whisper, avelin-stt), text-to-speech (tts-1), and image generation

#### [Model Families Overview](models/families.md)
Deep dive into each model family's positioning, capabilities, and use cases.

---

## 🛠️ Developer Resources

Build applications with the AVELIN API. Compatible with OpenAI and Anthropic SDKs.

### API Documentation

- **[API Reference](api/index.md)** — Complete endpoint documentation for chat completions, embeddings, audio, and images
- **[Web Intelligence API](api/web-scraping.md)** — Firecrawl-compatible scrape, search, map, crawl, and extract on your AVELIN key
- **[API Quickstart](api/quickstart.md)** — Step-by-step guide to your first API call
- **[SDKs & Tools](api/sdks.md)** — Official SDKs, CLI tools, and editor integrations

### Integration Guides

Practical guides for integrating AVELIN with popular tools. We support 40+ OpenRouter-compatible apps.

#### Featured Integrations

- **[Browser Use](guides/browser-use.md)** — Browser automation (92K GitHub stars)
- **[Continue](guides/continue.md)** — Open-source VS Code/JetBrains assistant
- **[NanoClaw](guides/nanoclaw.md)** — Lightweight AI code analysis tool
- **[VT Code](guides/vt-code.md)** — Terminal-based AI coding assistant

#### Coding & IDE Tools

- **[Cursor](guides/coding/cursor.md)** — AI-first code editor
- **[Windsurf](guides/coding/windsurf.md)** — Codeium's AI-first IDE
- **[GitHub Copilot](guides/coding/github-copilot.md)** — BYOK Agent mode
- **[Cline](guides/coding/cline.md)** — VS Code autonomous coding agent
- **[Roo Code](guides/coding/roo-code.md)** — Autonomous VS Code coding agent
- **[Aider](guides/coding/aider.md)** — AI pair programming in terminal
- **[Claude Code](guides/coding/claude-code.md)** — Anthropic's CLI coding tool
- **[Codex](guides/coding/codex.md)** — OpenAI's terminal AI assistant
- **[OpenCode](guides/coding/opencode.md)** — Terminal-based coding assistant
- **[Kilo CLI](guides/coding/kilo-cli.md)** — Terminal AI coding tool
- **[More coding tools →](guides/coding/README.md)** — nanocode, Autohand, GitBug, Shakespeare, and more

#### Chat & Desktop Clients

- **[LibreChat](guides/chat/librechat.md)** — Self-hosted ChatGPT-style interface
- **[Chatbox](guides/chat/chatbox.md)** — Cross-platform desktop client
- **[BoltAI](guides/chat/boltai.md)** — Native macOS chat app
- **[SillyTavern](guides/chat/sillytavern.md)** — Power-user chat frontend
- **[More chat clients →](guides/chat/README.md)** — ChatLima, Chorus, Warden, Skales

#### Automation & Productivity

- **[Hermes Agent](guides/hermes.md)** — AI agent framework (recommended)
- **[OpenClaw](guides/coding/openclaw.md)** — Multi-agent orchestration
- **[More automation tools →](guides/automation/README.md)** — Agent Zero, AiAssistWorks, Ottex, and more

#### Creative Tools

- **[Creative tools →](guides/creative/README.md)** — Novelcrafter, Aventura, Project AIRI

#### Frameworks & SDKs

- **[Frameworks →](guides/frameworks/README.md)** — Mastra, openrouter-rs, OpenRouter MCP

#### Observability

- **[Observability tools →](guides/observability/README.md)** — Helicone, analystOS, Cloudflare AI Gateway, Maxim AI, PostHog

---

## 💼 Business Use Cases

See how organizations use AVELIN to transform workflows.

### [Use Cases Overview](use-cases/index.md)
Real-world scenarios and implementation patterns.

- **[Sales Intelligence](use-cases/sales-account-intelligence.md)** — Account research, outreach personalization, and deal intelligence
- **[Service Support](use-cases/service-support-resolution.md)** — Automated ticket resolution, knowledge base search, and customer support
- **[Executive Operations](use-cases/executive-operations-briefing.md)** — Meeting prep, decision support, and strategic analysis
- **[Personal Productivity](use-cases/personal-productivity-automation.md)** — Email drafting, task management, and workflow automation

---

## 🏢 Platform Architecture

Understand how AVELIN works under the hood.

### Core Components

- **[Platform Architecture](architecture.md)** — System design, routing logic, and data flow
- **[How It Works](systems/how-it-works.md)** — Technical overview of the model orchestration layer

### Platform Modules

- **[AVELIN-API](systems/avelin-api.md)** — The governed API layer with automatic failover and retry logic
- **[AVELIN-GPT](systems/avelin-conversational-interface.md)** — Conversational interface for non-technical users
- **[AVELIN-MCP](systems/mcp-integration-platform.md)** — Model Context Protocol for tool integrations
- **[Y-RAY](systems/document-intelligence-rag.md)** — Document search and RAG (retrieval-augmented generation)

### MCP Integrations

Connect AVELIN to your productivity tools via MCP (Model Context Protocol).

- **[MCP Overview](mcp/index.md)** — What MCP is and how it works
- **[Google Calendar](mcp/google-calendar.md)** — Create, update, and query calendar events
- **[Gmail](mcp/gmail.md)** — Send emails, search messages, manage drafts
- **[Google Tasks](mcp/google-tasks.md)** — Create and manage task lists
- **[Google Contacts](mcp/google-contacts.md)** — Search and manage contact information
- **[Cross-Service Workflows](mcp/cross-service-scenarios.md)** — Multi-tool automation examples
- **[Action Reference](mcp/action-reference.md)** — Complete list of MCP actions and parameters

---

## 📊 Benefits & Value

Why organizations choose AVELIN over direct provider APIs.

### [Benefits Overview](benefits/index.md)
Strategic, operational, and competitive advantages.

- **[Strategic Benefits](benefits/platform-strategic-benefits.md)** — Vendor neutrality, data sovereignty, and governance
- **[Operational Benefits](benefits/operational-benefits.md)** — Cost optimization, reliability, and simplified operations
- **[Competitive Advantages](benefits/competitive-advantages.md)** — How AVELIN compares to direct provider relationships
- **[Evaluation](benefits/benchmark-results.md)** — How AVELIN models are evaluated; results at eval.avelin.ai

---

## 🔒 Security & Operations

Enterprise-grade security, reliability, and operational control.

### [Operations Overview](operations/index.md)
Deployment, administration, and governance guidance.

- **[Security & Governance](operations/security-reliability-governance.md)** — Encryption, access control, and compliance
- **[Deployment & Administration](operations/deployment-and-administration.md)** — Installation, configuration, and management
- **[Troubleshooting & KPIs](operations/troubleshooting-and-kpis.md)** — Common issues, monitoring, and performance metrics

---

## 📚 Reference

### [Glossary](glossary.md)
Definitions of key terms and concepts used throughout the documentation.

---

## Need Help?

- **Getting started**: Read the [API Quickstart](api/quickstart.md)
- **Choosing a model**: See the [Model Catalog](models/README.md)
- **Integration questions**: Check [SDKs & Tools](api/sdks.md) or [Guides](guides/hermes.md)
- **Enterprise deployment**: Review [Operations](operations/index.md)

---

## Documentation Structure

```
docs/
├── api/              # API reference, quickstart, SDKs
├── models/           # Model catalog and family overviews
├── guides/           # Integration guides (Hermes, etc.)
├── use-cases/        # Business scenarios and workflows
├── systems/          # Platform modules and architecture
├── mcp/              # MCP integrations and actions
├── benefits/         # Value proposition and benchmarks
├── operations/       # Deployment, security, troubleshooting
└── README.md         # This file
```

---

URL: https://avelin.ai/docs/api

# AVELIN API Reference

AVELIN is an **AI Laboratory** that builds Cross-Model MoE (Mixture-of-Experts) models — trained AI systems that fuse knowledge from multiple model architectures into single unified models. We don't just route; we build and fuse. The AVELIN API serves these models through a single governed endpoint, delivering superior results at every tier.

**Enterprise licensing available:** Deploy AVELIN models on your own GPUs with full data sovereignty. [Request a POC](/poc#request) for enterprise licensing options.

The API is a **drop-in replacement** for both the OpenAI and Anthropic wire
formats — most existing SDKs and tools work by changing only **2 lines of code**:
the base URL and API key.

Supported workloads include chat, reasoning, agentic tool use, coding,
embedding, transcription, and image generation — with automatic model selection,
failover, and unified billing behind one API key.

- New here? Start with the [Quickstart](quickstart.md).
- Using an SDK or CLI? See [SDKs & Tools](sdks.md).
- Choosing a model? See the [Model Catalog](../models/README.md).
- Need live web data? See the [Web Intelligence API](web-scraping.md) — a Firecrawl-compatible scrape/search/crawl surface.

---

## Base URL

```
https://api.avelin.ai
```

Model endpoints are served under the `/v1` path prefix. The Firecrawl-compatible
[Web Intelligence API](web-scraping.md) is served under `/v2`.

## Authentication

Every request requires a bearer token:

```
Authorization: Bearer sk-avelin-YOUR_KEY
```

Keys are issued and managed from your AVELIN account. Treat keys as secrets —
never embed them in client-side code or commit them to source control. Use
separate keys per application or environment so they can be rotated and
spend-tracked independently.

## Compatibility surfaces

| Surface | Endpoint | Best for |
|---|---|---|
| **OpenAI-compatible** | `/v1/chat/completions` | OpenAI SDKs, most third-party tools |
| **Anthropic-compatible** | `/v1/messages` | Anthropic SDKs, Claude-style clients, fine-grained thinking control |

Both surfaces are backed by the same models and return equivalent results —
pick whichever matches your existing stack. Switching is a **2-line change**:
update the base URL and API key.

For live web data — scrape, search, map, crawl, and extract — AVELIN also offers a
**Firecrawl-compatible** surface under `/v2`, usable from the official Firecrawl SDK with
your AVELIN key. See the [Web Intelligence API](web-scraping.md).

---

## Endpoints

### `POST /v1/chat/completions` (OpenAI-compatible)

Create a chat completion.

**Request**

```bash
curl https://api.avelin.ai/v1/chat/completions \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "avelin-pro",
    "messages": [
      { "role": "system", "content": "You are a helpful assistant." },
      { "role": "user", "content": "Summarize the benefits of vendor-neutral AI." }
    ],
    "max_tokens": 1024
  }'
```

**Common parameters**

| Field | Type | Description |
|---|---|---|
| `model` | string | A public model alias, e.g. `avelin-pro` (see [Model Catalog](../models/README.md)). |
| `messages` | array | Conversation messages with `role` (`system` / `user` / `assistant`) and `content`. |
| `max_tokens` | integer | Maximum tokens to generate. |
| `stream` | boolean | `true` to stream tokens as Server-Sent Events. |
| `tools` | array | Tool/function definitions (OpenAI function-calling format). |
| `tool_choice` | string | `"auto"` or `"none"`. |
| `reasoning_effort` | string | `"high"` to request deeper reasoning on models that support it. |

**Response**

```json
{
  "id": "chatcmpl-...",
  "object": "chat.completion",
  "model": "avelin-pro",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Vendor-neutral AI reduces lock-in by...",
        "reasoning_content": "The user wants a concise summary..."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 24,
    "completion_tokens": 180,
    "total_tokens": 204
  }
}
```

On reasoning-capable models, the model's thinking is returned in the
`reasoning_content` field, separate from the final `content`.

---

### `POST /v1/messages` (Anthropic-compatible)

Create a message with explicit thinking control.

**Request**

```bash
curl https://api.avelin.ai/v1/messages \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "avelin-coding-fast",
    "max_tokens": 1024,
    "system": "You are a senior software engineer.",
    "messages": [
      { "role": "user", "content": "Write a Python function to debounce calls." }
    ]
  }'
```

**Common parameters**

| Field | Type | Description |
|---|---|---|
| `model` | string | A public model alias. |
| `max_tokens` | integer | Maximum tokens to generate (required). |
| `messages` | array | Messages with `role` and `content`. |
| `system` | string | System prompt. |
| `thinking` | object | `{ "type": "disabled" }` to turn off extended thinking; enabled by default where supported. |
| `tools` | array | Tool definitions (Anthropic `input_schema` format). |
| `stream` | boolean | `true` for streaming. |

**Response**

```json
{
  "id": "resp_...",
  "type": "message",
  "role": "assistant",
  "model": "avelin-coding-fast",
  "content": [
    { "type": "text", "text": "Here is a debounce helper..." }
  ],
  "stop_reason": "end_turn",
  "usage": { "input_tokens": 18, "output_tokens": 210, "total_tokens": 228 }
}
```

---

### `GET /v1/models`

List the models available to your key.

```bash
curl https://api.avelin.ai/v1/models \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY"
```

```json
{
  "object": "list",
  "data": [
    { "id": "avelin-ultra", "object": "model", "owned_by": "avelin" },
    { "id": "avelin-pro",   "object": "model", "owned_by": "avelin" },
    { "id": "avelin-fast",  "object": "model", "owned_by": "avelin" }
  ]
}
```

See the [Model Catalog](../models/README.md) for descriptions and capabilities.

---

### `POST /v1/embeddings`

Generate vector embeddings (model: `bge-m3`).

```bash
curl https://api.avelin.ai/v1/embeddings \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bge-m3",
    "input": "AVELIN is a sovereign AI platform."
  }'
```

```json
{
  "object": "list",
  "data": [
    { "object": "embedding", "index": 0, "embedding": [0.0123, -0.0456, "..."] }
  ],
  "model": "bge-m3",
  "usage": { "prompt_tokens": 9, "total_tokens": 9 }
}
```

---

### `POST /v1/audio/transcriptions`

Transcribe audio to text (models: `whisper-large-v3`, `whisper-large-v3-turbo`).

```bash
curl https://api.avelin.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -F model="whisper-large-v3-turbo" \
  -F file="@meeting.mp3"
```

---

### `POST /v1/images/generations`

Generate images (models: `avelin-imagegen`, `avelin-imagegen-pro`).

```bash
curl https://api.avelin.ai/v1/images/generations \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "avelin-imagegen",
    "prompt": "A minimalist logo for a sovereign AI platform"
  }'
```

---

## Streaming

Set `stream: true` to receive tokens incrementally.

**OpenAI format** (`/v1/chat/completions`) — newline-delimited `data:` lines:

```
data: {"id":"...","object":"chat.completion.chunk","model":"avelin-pro","choices":[{"index":0,"delta":{"reasoning_content":"The"}}]}
data: {"id":"...","object":"chat.completion.chunk","model":"avelin-pro","choices":[{"index":0,"delta":{"content":"Vendor"}}]}
data: [DONE]
```

**Anthropic format** (`/v1/messages`) — Server-Sent Events with `event:` prefixes:

```
event: content_block_start
data: {"type":"content_block_start","index":0,"content_block":{"type":"thinking","thinking":""}}

event: content_block_delta
data: {"type":"content_block_delta","index":1,"delta":{"type":"text_delta","text":"Hi"}}
```

Thinking/reasoning content streams separately from the final answer text.

---

## Tool & function calling

Both surfaces support tool calling. OpenAI format:

```json
{
  "model": "avelin-agentic-pro",
  "messages": [{ "role": "user", "content": "What's the weather in Abu Dhabi?" }],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get current weather for a city",
        "parameters": {
          "type": "object",
          "properties": { "city": { "type": "string" } },
          "required": ["city"]
        }
      }
    }
  ],
  "tool_choice": "auto"
}
```

The Anthropic surface uses the equivalent `tools` array with `input_schema`.

---

## Reasoning & thinking control

| Goal | OpenAI surface | Anthropic surface |
|---|---|---|
| Request deeper reasoning | `"reasoning_effort": "high"` | enabled by default on supported models |
| Read the model's thinking | `reasoning_content` field in the response | separate `thinking` content block |
| Disable thinking | not applicable (reasoning enabled by default on supported models) | `"thinking": { "type": "disabled" }` |

---

## Errors

Errors follow the OpenAI error shape:

```json
{ "error": { "message": "Invalid API key", "type": "authentication_error", "code": 401 } }
```

| Status | Meaning |
|---|---|
| `400` | Malformed request (bad JSON, missing required field). |
| `401` | Missing or invalid API key. |
| `404` | Unknown model alias or endpoint. |
| `429` | Rate or budget limit reached — back off and retry. |
| `5xx` | Transient inference issue — the platform retries and fails over automatically; retry with backoff. |

---

## Notes & behavior

- Reasoning/thinking is enabled by default on models that support it.
- The platform automatically retries and fails over on transient inference
  errors, so occasional latency spikes are smoothed for you.
- `max_tokens` is respected; responses report `finish_reason: "length"` when the
  limit is reached.
- Pick the model by **capability tier** (see the [Model Catalog](../models/README.md));
  AVELIN's Cross-Model MoE models handle inference with fused intelligence for you.

## Related

- [Quickstart](quickstart.md)
- [SDKs & Tools](sdks.md)
- [Model Catalog](../models/README.md)
- [AVELIN-API platform overview](../systems/avelin-api.md)

---

URL: https://avelin.ai/docs/api/quickstart

# API Quickstart

Make your first AVELIN API call in under five minutes. AVELIN is an **AI Laboratory** that builds Cross-Model MoE (Mixture-of-Experts) models — trained AI systems that fuse knowledge from multiple model architectures into single unified models. We don't just route; we build and fuse. The API serves these models through a single governed endpoint.

**Enterprise licensing available:** Deploy AVELIN models on your own GPUs with full data sovereignty. [Request a POC](/poc#request) for enterprise licensing options.

The API is a **drop-in replacement**: change only **2 lines of code** (base URL and API key) to migrate from OpenAI or Anthropic.

## 1. Get an API key

Create a key from your AVELIN account. It will look like:

```
sk-ave...xxxx
```

Keep it secret. Set it as an environment variable so it never lands in code:

```bash
export AVELIN_API_KEY="sk-ave...xxxx"
```

---

## 2. Choose your track

| Track | When to use |
|---|---|
| [**OpenAI-compatible**](#openai-track) | You already use the OpenAI SDK, or your tool expects `/v1/chat/completions` |
| [**Anthropic-compatible**](#anthropic-track) | You already use the Anthropic SDK, or you want fine-grained thinking control via `/v1/messages` |

Both tracks are backed by the same models — pick whichever matches your stack.

---

## OpenAI track {#openai-track}

**Base URL:** `https://api.avelin.ai/v1`  
**Endpoint:** `POST /v1/chat/completions`

### cURL

```bash
curl https://api.avelin.ai/v1/chat/completions \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "avelin-pro",
    "messages": [
      { "role": "user", "content": "Give me three productivity tips." }
    ]
  }'
```

### Python (OpenAI SDK)

```python
from openai import OpenAI

client = OpenAI(
    base_url="https://api.avelin.ai/v1",
    api_key="sk-ave...xxxx",  # or os.environ["AVELIN_API_KEY"]
)

resp = client.chat.completions.create(
    model="avelin-pro",
    messages=[{"role": "user", "content": "Give me three productivity tips."}],
)
print(resp.choices[0].message.content)
```

### JavaScript / TypeScript (OpenAI SDK)

```ts
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.avelin.ai/v1",
  apiKey: process.env.AVELIN_API_KEY,
});

const resp = await client.chat.completions.create({
  model: "avelin-pro",
  messages: [{ role: "user", content: "Give me three productivity tips." }],
});
console.log(resp.choices[0].message.content);
```

### Response shape (OpenAI)

```json
{
  "id": "chatcmpl-...",
  "object": "chat.completion",
  "model": "avelin-pro",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "1. Time-block your calendar...",
        "reasoning_content": "The user wants practical tips..."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 12,
    "completion_tokens": 180,
    "total_tokens": 192
  }
}
```

### Streaming (OpenAI)

```python
stream = client.chat.completions.create(
    model="avelin-pro",
    messages=[{"role": "user", "content": "Write a haiku about sovereignty."}],
    stream=True,
)
for chunk in stream:
    delta = chunk.choices[0].delta
    if delta.content:
        print(delta.content, end="", flush=True)
```

---

## Anthropic track {#anthropic-track}

**Base URL:** `https://api.avelin.ai`  
**Endpoint:** `POST /v1/messages`  
**Header:** `anthropic-version: 2023-06-01`

### cURL

```bash
curl https://api.avelin.ai/v1/messages \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "avelin-pro",
    "max_tokens": 1024,
    "messages": [
      { "role": "user", "content": "Give me three productivity tips." }
    ]
  }'
```

### Python (Anthropic SDK)

```python
from anthropic import Anthropic

client = Anthropic(
    base_url="https://api.avelin.ai",
    api_key="sk-ave...xxxx",  # or os.environ["AVELIN_API_KEY"]
)

msg = client.messages.create(
    model="avelin-pro",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Give me three productivity tips."}],
)
print(msg.content[0].text)
```

### JavaScript / TypeScript (Anthropic SDK)

```ts
import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  baseURL: "https://api.avelin.ai",
  apiKey: process.env.AVELIN_API_KEY,
});

const msg = await client.messages.create({
  model: "avelin-pro",
  max_tokens: 1024,
  messages: [{ role: "user", content: "Give me three productivity tips." }],
});
console.log(msg.content[0].text);
```

### Response shape (Anthropic)

```json
{
  "id": "resp_...",
  "type": "message",
  "role": "assistant",
  "model": "avelin-pro",
  "content": [
    { "type": "text", "text": "1. Time-block your calendar..." }
  ],
  "stop_reason": "end_turn",
  "usage": { "input_tokens": 12, "output_tokens": 180, "total_tokens": 192 }
}
```

### System prompts (Anthropic)

The Anthropic surface uses a top-level `system` parameter instead of a system
message in the messages array:

```python
msg = client.messages.create(
    model="avelin-coding-fast",
    max_tokens=1024,
    system="You are a senior software engineer.",
    messages=[{"role": "user", "content": "Write a debounce decorator."}],
)
```

### Thinking control (Anthropic)

Extended thinking is enabled by default on reasoning-capable models. To disable:

```python
msg = client.messages.create(
    model="avelin-pro",
    max_tokens=1024,
    thinking={"type": "disabled"},
    messages=[{"role": "user", "content": "Summarize this document."}],
)
```

When thinking is enabled, the response includes a separate `thinking` content
block before the final `text` block.

### Streaming (Anthropic)

```python
with client.messages.stream(
    model="avelin-pro",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Write a haiku about sovereignty."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)
```

---

## 3. Pick the right model

All models are built with AVELIN's Cross-Model MoE technology in our AI Lab.

| If you need… | Use |
| --- | --- |
| Fast, low-cost everyday responses | `avelin-fast` |
| Balanced speed and quality | `avelin-pro` |
| Deepest reasoning for high-stakes work | `avelin-ultra` |
| Tool-using agent workflows | `avelin-agentic-pro` / `avelin-agentic-ultra` |
| Software engineering | `avelin-coding-fast` / `avelin-coding-pro` / `avelin-coding-ultra` |

**Cost savings vs. direct providers:**
- Intelligence family: **up to 85% cheaper**
- Coding family: **up to 86% cheaper**
- Agentic family: **up to 87% cheaper**

Full descriptions: [Model Catalog](../models/README.md).

## Key differences between tracks

| | OpenAI track | Anthropic track |
|---|---|---|
| **Base URL** | `https://api.avelin.ai/v1` | `https://api.avelin.ai` |
| **Endpoint** | `/v1/chat/completions` | `/v1/messages` |
| **Auth header** | `Authorization: Bearer sk-avelin-YOUR_KEY` | `Authorization: Bearer sk-avelin-YOUR_KEY` |
| **Extra header** | — | `anthropic-version: 2023-06-01` |
| **`max_tokens`** | Optional | **Required** |
| **System prompt** | Message with `"role": "system"` | Top-level `"system"` parameter |
| **Thinking control** | `"reasoning_effort": "high"` | `"thinking": { "type": "disabled" }` |
| **Thinking output** | `reasoning_content` field | Separate `thinking` content block |
| **Tool format** | OpenAI function-calling | Anthropic `input_schema` |

## Next steps

- Full [API Reference](index.md) — every endpoint, parameter, and response shape
- [SDKs & Tools](sdks.md) — CLI and editor integration configuration
- [Model Catalog](../models/README.md) — capabilities and selection guidance

---

URL: https://avelin.ai/docs/models

# AVELIN Model Catalog

AVELIN is an **AI Laboratory** that builds **Cross-Model MoE** — a trained Mixture-of-Experts meta-model architecture with difficulty-based cascading that fuses open-weight models into **9 specialist families across 3 categories**: Intelligence, Coding, and Agentic.

This is not routing or pass-through. Cross-Model MoE is a **trained meta-model, not a router** — with learned gating and specialist weighting that synthesizes intelligence exceeding any single model. A prompt-vector classifier (trained on 1 year of production data, 90% accuracy) applies **difficulty-based cascading** and **cost-aware routing** across specialist layers. Routers pick one model via if/else rules. We cascade across trained experts and synthesize. New models are integrated within **48 hours** of release, ensuring the platform is continuously self-improving.

All models are available via **enterprise licensing** (deploy on your GPUs — [request a POC](/poc#request)) or **cloud API** (usage-based, $0.285–$12.50 per 1M tokens). See the [API Reference](../api/reference.md) to start calling them.

Every family follows the same three-tier naming: **fast** (cost-efficient), **pro** (balanced), **ultra** (premium). Five models carried older names before this scheme — see [Legacy Model Names](#legacy-model-names) below; the old names keep working.

**New to AVELIN models?** Start with the [Model Families Overview](families.md) for detailed descriptions of each family's capabilities and use cases.

---

## Intelligence Family

**Frontier reasoning and multimodal intelligence — up to 85% cheaper than frontier alternatives.**

The Intelligence family delivers deep reasoning, strategic analysis, and multimodal understanding for the most demanding enterprise workloads. All models feature **256K token context windows** — enough to process entire documents, codebases, or conversation histories in a single request — and **native vision support**.

| Model | Best For | Context Window | Max Output | Input/Output $/M |
| ------- | ---------- | ---------------- | ------------ | ------------------- |
| [**avelin-ultra**](avelin-ultra.md) | Deep research, strategic analysis, complex decisions | **256K** | 65K | $2.50 / $12.50 |
| [**avelin-pro**](avelin-pro.md) | Daily workflows, multimodal tasks, production economics | **256K** | 65K | $1.40 / $4.40 |
| [**avelin-fast**](avelin-fast.md) | Real-time chat, classification, high-volume workloads | **256K** | 65K | $0.30 / $1.20 |

**Why 256K context transforms workflows:** Analyze entire legal contracts, process full research papers, or maintain context across days of conversation — all without chunking or summarization. This is intelligence that sees the complete picture.

---

## Coding Family

**Enterprise coding agents — up to 86% cheaper than frontier alternatives.**

The Coding family is purpose-built for real software engineering work. All coding models feature **1M token context windows** — load entire repositories and understand cross-file dependencies without fragmentation.

| Model | Best For | Context Window | Max Output | Input/Output $/M |
| ------- | ---------- | ---------------- | ------------ | ------------------- |
| [**avelin-coding-ultra**](avelin-coding-ultra.md) | System design, security review, complex refactors | **1M** | 65K | $2.50 / $12.50 |
| [**avelin-coding-pro**](avelin-coding-pro.md) | Production workloads, CI/CD | **1M** | 65K | $1.40 / $4.40 |
| [**avelin-coding-fast**](avelin-coding-fast.md) | Daily coding workflows, debugging | **1M** | 65K | $0.30 / $1.20 |

**The 1M context advantage:** This isn't just a large context window — it's whole-repository intelligence. Load your entire codebase, trace execution paths across modules, understand cross-file dependencies, and refactor with complete awareness. No more "I can only see 5 files at a time" limitations.

---

## Agentic Family

**Purpose-built for autonomous agents — up to 87% cheaper than frontier alternatives.**

The Agentic family is optimized for long-horizon agentic loops, tool calling, and parallel orchestration. All agentic models feature **256K token context windows** — enough for agents to maintain coherent state across hundreds of steps without losing track of goals, at economics tuned for many calls per workflow.

| Model | Best For | Context Window | Max Output | Input/Output $/M |
| ------- | ---------- | ---------------- | ------------ | ------------------- |
| [**avelin-agentic-ultra**](avelin-agentic-ultra.md) | Complex multi-step planning, security workflows | **256K** | 65K | $2.38 / $11.88 |
| [**avelin-agentic-pro**](avelin-agentic-pro.md) | Tool-calling agents, regulated industries | **256K** | 65K | $1.26 / $3.96 |
| [**avelin-agentic-fast**](avelin-agentic-fast.md) | High-volume agent swarms, bulk operations | **256K** | 65K | $0.285 / $1.14 |

**Why context size matters for agents:** Autonomous workflows require sustained reasoning across hundreds of tool calls, API interactions, and decision points. With 256K context, your agents can maintain coherent state, remember complex goals, and execute multi-step plans without fragmentation or context loss.

---

## Legacy Model Names

Five models were renamed when the lineup adopted the uniform fast/pro/ultra scheme. The old names keep working indefinitely as server-side aliases — no client change required, billing follows the canonical tier. New integrations should use the canonical names.

| Legacy name | Canonical name |
|-------------|----------------|
| `avelin-coding` | [`avelin-coding-fast`](avelin-coding-fast.md) |
| `avelin-coding-plus` | [`avelin-coding-pro`](avelin-coding-pro.md) |
| `avelin-coding-architect` | [`avelin-coding-ultra`](avelin-coding-ultra.md) |
| `avelin-agentic` | [`avelin-agentic-pro`](avelin-agentic-pro.md) |
| `avelin-agentic-high` | [`avelin-agentic-ultra`](avelin-agentic-ultra.md) |

---

## Utility Models

Specialized models for specific tasks beyond conversational AI.

| Model | Type | Use Case |
|-------|------|----------|
| **bge-m3** | Embeddings | Semantic search, RAG, clustering |
| **whisper-large-v3** | Transcription | Maximum accuracy, batch processing |
| **whisper-large-v3-turbo** | Transcription | Real-time, high volume |
| **avelin-stt** | Transcription | Streaming-friendly speech-to-text |
| **tts-1** | Text-to-Speech | Low latency, cost-efficient voice |
| **tts-1-hd** | Text-to-Speech | Highest audio quality |
| **avelin-imagegen** | Image generation | Quick iterations, prototypes |
| **avelin-imagegen-pro** | Image generation | Production assets, high quality |

See [Utility Models](utility-models.md) for detailed API documentation.

---

## Quick Selection Guide

| Use Case | Recommended Model | Alternative |
|----------|-------------------|-------------|
| **Strategic analysis, executive decisions** | avelin-ultra | avelin-pro |
| **Daily business workflows** | avelin-pro | avelin-fast |
| **High-volume chat, classification** | avelin-fast | avelin-pro |
| **Software development** | avelin-coding-fast | avelin-coding-pro |
| **High-throughput coding pipelines** | avelin-coding-pro | avelin-coding-fast |
| **System architecture, security review** | avelin-coding-ultra | avelin-coding-pro |
| **Tool-calling agents** | avelin-agentic-pro | avelin-agentic-fast |
| **Complex agentic workflows** | avelin-agentic-ultra | avelin-agentic-pro |
| **High-volume agent swarms** | avelin-agentic-fast | avelin-agentic-pro |
| **Semantic search, RAG** | bge-m3 | — |
| **Audio transcription** | whisper-large-v3-turbo | whisper-large-v3 |
| **Text-to-speech** | tts-1 | tts-1-hd |
| **Image generation** | avelin-imagegen | avelin-imagegen-pro |

---

## Model Selection Strategy

For the best balance of quality, speed, and cost:

1. **Start with cost-efficient tiers** — Use `avelin-fast`, `avelin-agentic-fast`, or `avelin-coding-fast` for routine work
2. **Escalate when needed** — Use `avelin-ultra`, `avelin-coding-ultra`, or `avelin-agentic-ultra` for complex, high-stakes tasks
3. **Use specialized families** — Coding models for software engineering, agentic models for tool-calling workflows
4. **Leverage Cross-Model MoE** — AVELIN's trained meta-model automatically selects the optimal specialist pathway for your request

### Cost Optimization

- **68% of requests** can be handled by cost-efficient tiers (fast models)
- **20% of requests** benefit from balanced tiers (pro models)
- **12% of requests** require premium tiers (ultra models)

This difficulty-based cascading delivers **42% cost reduction** vs always-premium routing while maintaining quality. Combined with our Cross-Model MoE pricing advantage (up to 87% cheaper than frontier), enterprises see dramatic savings from day one — whether deployed via cloud API or on their own GPUs under enterprise license.

---

## Model Capabilities Matrix

| Model | Context | Max Output | Reasoning | Vision | Function Calling | Streaming | System Messages | Prompt Caching |
|-------|---------|------------|-----------|--------|------------------|-----------|-----------------|----------------|
| avelin-ultra | **256K** | 65K | ✅ Full | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-pro | **256K** | 65K | ✅ Full | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-fast | **256K** | 65K | ⚠️ Off by default | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-coding-fast | **1M** | 65K | ⚠️ Off by default | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-coding-pro | **1M** | 65K | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-coding-ultra | **1M** | 65K | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-agentic-fast | **256K** | 65K | ⚠️ Off by default | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-agentic-pro | **256K** | 65K | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-agentic-ultra | **256K** | 65K | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |

**Prompt Caching:** Cuts input costs by roughly 80% on cache hits for pro and ultra, and by 90% on all three fast tiers (cache write is 1.25× input; cache read is ~5x cheaper than fresh input on pro/ultra and 10x on the fast tiers). Supported on models with ✅ Yes. See [Pricing](../pricing.md) for per-tier cache rates.

**Context Window Guide:**
- **1M tokens:** Whole-repository intelligence — load entire codebases, process massive documents, maintain state across 100+ step agent workflows
- **256K tokens:** Extended context — analyze full research papers, legal contracts, or days of conversation history

---

---

## Pricing

See [Pricing](../pricing.md) for complete pricing information including enterprise licensing, cloud API rates, cost optimization strategies, and billing details.

---

## Related Documentation

- [Pricing](../pricing.md) - Complete pricing and billing information
- [Model Families Overview](families.md) - Detailed family descriptions
- [API Reference](../api/reference.md) - Complete API documentation
- [API Quickstart](../api/quickstart.md) - Get started in 5 minutes
- [How AVELIN Works](../systems/how-it-works.md) - Technical architecture
- [Evaluation](../benefits/benchmark-results.md) - How AVELIN models are evaluated
- [AVELIN-API Platform](../systems/avelin-api.md) - Platform overview

---

URL: https://avelin.ai/docs/models/families

# AVELIN Model Families

AVELIN is an **AI Laboratory** that builds and trains **Cross-Model MoE** — a proprietary Mixture-of-Experts meta-model architecture with difficulty-based cascading. We fuse open-weight models into **9 specialist families across 3 categories**, producing trained intelligence that exceeds any single model:

| Category | Families | Cost Advantage |
|----------|----------|----------------|
| **Intelligence** | fast, pro, ultra | up to 85% cheaper than frontier alternatives |
| **Coding** | coding-fast, coding-pro, coding-ultra | up to 86% cheaper than frontier alternatives |
| **Agentic** | agentic-fast, agentic-pro, agentic-ultra | up to 87% cheaper than frontier alternatives |

Cross-Model MoE is not routing or pass-through — it is a **trained meta-model, not a router**. A prompt-vector classifier (trained on 1 year of production data, 90% accuracy) classifies each request and applies **difficulty-based cascading** and **cost-aware routing** across specialist layers. Routers pick one model via if/else rules. We cascade across trained experts and synthesize. The platform is continuously self-improving: new models are evaluated and integrated within **48 hours** of release.

Our models are available via **enterprise licensing** (sovereign deployment on your GPUs — [request a POC](/poc#request)) or **cloud API** (usage-based, $0.285–$12.50 per 1M tokens).

> **Renamed models:** `avelin-coding`, `avelin-coding-plus`, `avelin-coding-architect`, `avelin-agentic`, and `avelin-agentic-high` are legacy names for `avelin-coding-fast`, `avelin-coding-pro`, `avelin-coding-ultra`, `avelin-agentic-pro`, and `avelin-agentic-ultra`. The old names keep working as server-side aliases. See the [Model Catalog](README.md#legacy-model-names).

---

## Intelligence Family

**Frontier reasoning and multimodal intelligence — up to 85% cheaper than frontier alternatives.**

The Intelligence family is built for analysis, strategy, knowledge work, research synthesis, and customer-facing assistants where answer quality, reasoning depth, and multimodal understanding come first. Through Cross-Model MoE, these models deliver frontier-tier intelligence under a single API with difficulty-based cascading for resilience and SLA-grade uptime.

### Models in this family

| Model | Positioning | Context Window | Max Output | Input/Output $/M |
| ------- | ------------- | ---------------- | ------------ | ------------------- |
| [avelin-ultra](avelin-ultra.md) | Flagship reasoning | **256K tokens** | 65K | $2.50 / $12.50 |
| [avelin-pro](avelin-pro.md) | High-efficiency reasoning | **256K tokens** | 65K | $1.40 / $4.40 |
| [avelin-fast](avelin-fast.md) | Low-latency, high-throughput | **256K tokens** | 65K | $0.30 / $1.20 |

**Why 256K context matters:** Process entire codebases, lengthy documents, or extended conversations in a single request. No chunking, no context loss, no compromises.

### Key Capabilities

- **Deep multi-step reasoning** for complex problem solving
- **Multimodal understanding** (text, images, charts, documents)
- **Vision support on all three tiers** for image analysis and document understanding
- **System messages** for custom behavior and context
- **Function calling** for structured outputs and tool integration
- **Streaming** for real-time applications

### When to use the Intelligence family

- Strategic planning and executive decision support
- Complex document analysis and synthesis
- Customer-facing assistants requiring high quality
- Research and knowledge work
- Multimodal tasks involving images, charts, or documents
- Tasks where accuracy and reasoning depth are critical

### Pricing

See [Pricing](../pricing.md) for complete pricing information including enterprise licensing and cloud API rates.

---

## Coding Family

**Enterprise coding agents — up to 86% cheaper than frontier alternatives.**

The Coding family is purpose-built for software development workflows — from live debugging and root-cause analysis to multi-file refactors and whole-system design. These models are optimized specifically for software engineering, with tool use throughout; thinking is on by default on the pro and ultra tiers and off by default on the fast tier for latency. They're built for IDE plugins, autonomous coding agents, code-review pipelines, and engineering productivity platforms.

### Models in this family

| Model | Positioning | Context Window | Max Output | Throughput | Input/Output $/M |
| ------- | ------------- | ---------------- | ------------ | ------------ | ------------------- |
| [avelin-coding-fast](avelin-coding-fast.md) | Balanced engineering | **1M tokens** | 65K | ~50 tps | $0.30 / $1.20 |
| [avelin-coding-pro](avelin-coding-pro.md) | High-throughput | **1M tokens** | 65K | ~100 tps | $1.40 / $4.40 |
| [avelin-coding-ultra](avelin-coding-ultra.md) | Deep architecture | **1M tokens** | 65K | ~50 tps | $2.50 / $12.50 |

**Why 1M context is a game-changer:** Load entire repositories into context. Understand cross-file dependencies, trace execution paths across modules, and refactor with complete codebase awareness. This isn't just a large context window — it's whole-repository intelligence.

### Key Capabilities

- **Whole-repository understanding** with 1M-token context windows
- **Multi-file refactoring** with cross-cutting impact analysis
- **Live debugging** and root-cause analysis
- **Code generation** across multiple languages and frameworks
- **Test generation** and code review
- **Security analysis** and vulnerability detection
- **Architecture design** and system planning
- **Tool use** for IDE integration and automation

### When to use the Coding family

- Software development and engineering tasks
- Code generation, refactoring, and optimization
- Bug fixing and debugging
- Code review and security analysis
- System architecture and design decisions
- IDE plugin development
- Autonomous coding agents
- CI/CD pipeline integration

### Pricing

See [Pricing](../pricing.md) for complete pricing information including enterprise licensing and cloud API rates.

---

## Agentic Family

**Purpose-built for autonomous agents — up to 87% cheaper than frontier alternatives.**

The Agentic family is optimized for cost-effective agent workflows that fire thousands of model calls per workflow. Through Cross-Model MoE, these models maintain frontier reasoning quality while keeping per-step economics low, making them ideal for agent teams, research pipelines, RAG systems, and orchestration layers. They support on-prem deployment for data sovereignty and air-gapped environments.

### Models in this family

| Model | Positioning | Context Window | Max Output | Input/Output $/M |
| ------- | ------------- | ---------------- | ------------ | ------------------- |
| [avelin-agentic-pro](avelin-agentic-pro.md) | Standard workhorse | **256K tokens** | 65K | $1.26 / $3.96 |
| [avelin-agentic-ultra](avelin-agentic-ultra.md) | Top-tier agentic | **256K tokens** | 65K | $2.38 / $11.88 |
| [avelin-agentic-fast](avelin-agentic-fast.md) | Cost-optimized | **256K tokens** | 65K | $0.285 / $1.14 |

**The agentic advantage:** Purpose-built for autonomous workflows that think, plan, and execute across hundreds of steps. With 256K context, your agents can maintain coherent state across complex, long-running tasks without losing track of goals or context.

### Key Capabilities

- **Long-horizon planning** across multiple steps and tools
- **Reliable tool calling** with structured outputs
- **Parallel task orchestration** for agent swarms
- **Configurable reasoning effort** for cost optimization
- **On-prem deployment** for regulated industries
- **High throughput** for volume workloads

### When to use the Agentic family

- Autonomous and semi-autonomous agents
- Multi-tool workflows requiring planning and execution
- Research pipelines with iterative refinement
- RAG systems with retrieval and synthesis
- Orchestration layers managing multiple agents
- High-volume classification and extraction
- Regulated industries requiring data sovereignty
- Cost-sensitive agent backends

### Pricing

See [Pricing](../pricing.md) for complete pricing information including enterprise licensing and cloud API rates.

---

## Utility Models

Specialized models for specific tasks beyond conversational AI.

### Embeddings

**bge-m3** - Multilingual embedding model optimized for retrieval tasks.

- **Dimensions**: 1024
- **Max tokens**: 8,192
- **Languages**: 100+ languages supported
- **Use cases**: Semantic search, clustering, RAG, similarity matching

See [Pricing](../pricing.md) for current pricing.

**Example:**
```bash
curl https://api.avelin.ai/v1/embeddings \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bge-m3",
    "input": "AVELIN is a sovereign AI platform."
  }'
```

### Transcription

**whisper-large-v3**, **whisper-large-v3-turbo**, and **avelin-stt** - Speech-to-text models.

| Model | Speed | Accuracy | Use Case |
|-------|-------|----------|----------|
| whisper-large-v3 | Standard | Highest | Maximum accuracy, batch processing |
| whisper-large-v3-turbo | Fast | High | Real-time transcription, high volume |
| avelin-stt | Fast | High | Streaming-friendly speech-to-text |

- **Languages**: 100+ languages with automatic detection
- **Formats**: mp3, mp4, mpeg, mpga, m4a, wav, webm (max 25MB)

See [Pricing](../pricing.md) for current pricing.

**Example:**
```bash
curl https://api.avelin.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -F model="whisper-large-v3-turbo" \
  -F file="@meeting.mp3"
```

### Text-to-Speech

**tts-1** and **tts-1-hd** - Voice synthesis with OpenAI-compatible model and voice names.

| Model | Latency | Quality | Use Case |
|-------|---------|---------|----------|
| tts-1 | Low | Standard | Real-time voice, cost-efficient narration |
| tts-1-hd | Standard | Highest | Production voice assets |

- **Voices**: OpenAI-compatible voice names (alloy, onyx, ...)
- **Formats**: mp3, opus, pcm

See [Pricing](../pricing.md) for current pricing.

**Example:**
```bash
curl https://api.avelin.ai/v1/audio/speech \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tts-1",
    "voice": "alloy",
    "input": "Welcome to AVELIN."
  }' --output speech.mp3
```

### Image Generation

**avelin-imagegen** and **avelin-imagegen-pro** - Text-to-image models.

| Model | Quality | Speed | Use Case |
|-------|---------|-------|----------|
| avelin-imagegen | Standard | Fast | Quick iterations, prototypes |
| avelin-imagegen-pro | High | Standard | Production assets, high quality |

See [Pricing](../pricing.md) for current pricing.

- **Sizes**: 256x256, 512x512, 1024x1024, 1024x1792, 1792x1024
- **Styles**: Vivid, natural
- **Quality**: Standard, HD (pro only)

**Example:**
```bash
curl https://api.avelin.ai/v1/images/generations \
  -H "Authorization: Bearer sk-avelin-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "avelin-imagegen",
    "prompt": "A minimalist logo for a sovereign AI platform"
  }'
```

---

## Choosing the Right Model

### By Use Case

| Use Case | Recommended Model | Alternative |
|----------|-------------------|-------------|
| Strategic analysis, executive decisions | avelin-ultra | avelin-pro |
| Daily business workflows | avelin-pro | avelin-fast |
| High-volume chat, classification | avelin-fast | avelin-pro |
| Software development | avelin-coding-fast | avelin-coding-pro |
| High-throughput coding pipelines | avelin-coding-pro | avelin-coding-fast |
| System architecture, security review | avelin-coding-ultra | avelin-coding-pro |
| Tool-calling agents | avelin-agentic-pro | avelin-agentic-fast |
| Complex agentic workflows | avelin-agentic-ultra | avelin-agentic-pro |
| High-volume agent swarms | avelin-agentic-fast | avelin-agentic-pro |
| Semantic search, RAG | bge-m3 | — |
| Audio transcription | whisper-large-v3-turbo | whisper-large-v3 |
| Text-to-speech | tts-1 | tts-1-hd |
| Image generation | avelin-imagegen | avelin-imagegen-pro |

### By Priority

**Quality first:** avelin-ultra, avelin-coding-ultra, avelin-agentic-ultra

**Cost first:** avelin-fast, avelin-agentic-fast, avelin-coding-fast

**Speed first:** avelin-fast, avelin-coding-pro, avelin-agentic-fast

**Balanced:** avelin-pro, avelin-coding-pro, avelin-agentic-pro

---

## Model Capabilities Matrix

| Model | Reasoning | Vision | Function Calling | Streaming | System Messages | Prompt Caching |
|-------|-----------|--------|------------------|-----------|-----------------|----------------|
| avelin-ultra | ✅ Full | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-pro | ✅ Full | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-fast | ⚠️ Limited | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-coding-fast | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-coding-pro | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-coding-ultra | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-agentic-fast | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-agentic-pro | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| avelin-agentic-ultra | ✅ Full | ❌ No | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |

---

---

## Related Documentation

- [Pricing](../pricing.md) - Complete pricing and billing information
- [API Reference](../api/reference.md) - Complete API documentation
- [Quickstart](../api/quickstart.md) - Get started in 5 minutes
- [How AVELIN Works](../systems/how-it-works.md) - Technical architecture
- [Evaluation](../benefits/benchmark-results.md) - How AVELIN models are evaluated

---

URL: https://avelin.ai/docs/systems

# Systems and Product Catalog

This section documents AVELIN as an **AI Laboratory** built on three pillars — **AI Lab** (primary), **Intelligence Delivery**, and **Enterprise Platform** — with Cross-Model MoE technology at its core.

AVELIN builds and fuses leading open models into 9 specialist families using trained Cross-Model Mixture-of-Experts technology, delivering enterprise-grade intelligence at significantly lower cost than direct provider access. The platform is self-improving, with new models integrated within 48 hours of release. AVELIN's primary deployment model is enterprise licensing.

## Platform Pillars

| Pillar | Focus |
|--------|-------|
| **AI Lab** (Primary) | Cross-Model MoE research, model training and fusion, specialist family development |
| **Intelligence Delivery** | High-throughput intelligence delivery with Smart Routing Economy and cost optimization |
| **Enterprise Platform** | Governed deployment, security, compliance, and operational control for organizations |

## Core Product Pages

- [`avelin-conversational-interface.md`](avelin-conversational-interface.md) - **AVELIN-GPT** — Conversational workspace (Enterprise Platform pillar)
- [`avelin-api.md`](avelin-api.md) - **AVELIN-API** — Intelligence orchestration core (Cloud Token Factory pillar)
- [`mcp-integration-platform.md`](mcp-integration-platform.md) - **AVELIN-MCP** — Action execution integration (Enterprise Platform pillar)
- [`document-intelligence-rag.md`](document-intelligence-rag.md) - **Y-RAY** — Corporate Brain research (AI Lab pillar)
- Model descriptions: [`../models/README.md`](../models/README.md)

## Platform Modules

- [`how-it-works.md`](how-it-works.md) - Technical architecture and Cross-Model MoE flow
- [`data-and-search-foundation.md`](data-and-search-foundation.md) - Sovereign data and intelligence memory foundation
- [`voice-and-web-intelligence.md`](voice-and-web-intelligence.md) - Voice, external intelligence, and interaction modes
- [`security-and-edge-layer.md`](security-and-edge-layer.md) - Secure access and perimeter control
- [`avelin-security-trust.md`](avelin-security-trust.md) - Enterprise trust and compliance controls
- [`operations-and-observability.md`](operations-and-observability.md) - Reliability, monitoring, and operational governance

## Reading Order

1. Start with [How AVELIN Works](how-it-works.md) to understand the Cross-Model MoE architecture.
2. Continue with AVELIN-GPT for user-facing value.
3. Review AVELIN-API for orchestration, specialist selection, and cost governance.
4. Explore AVELIN-MCP and Y-RAY for execution and intelligence depth.
5. Finish with trust, security, and operations modules for enterprise readiness.

---

URL: https://avelin.ai/docs/systems/how-it-works

# How AVELIN Works

## Technical Architecture Overview

AVELIN implements **API-level Mixture-of-Experts** — a specialist selection and fusion layer that automatically optimizes every request for cost, quality, and reliability. This page explains the technical flow from request to response.

## The 6-Stage Request Flow

```
User Request
    ↓
┌─────────────────────────────────────────────────────────┐
│ Stage 1: Semantic Router (50ms)                         │
│ - Classifies task type: code, reasoning, chat, etc.     │
│ - Embeds request using semantic transformer             │
│ - KNN classifier selects optimal model family           │
└─────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────┐
│ Stage 2: Cross-Model MoE Classifier                     │
│ - Considers: cost, latency, quality, availability      │
│ - Selects specific model tier (ultra/pro/fast)         │
│ - Can route to multiple models in parallel (fusion)    │
└─────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────┐
│ Stage 3: Model Fusion Layer (if enabled)                │
│ - Queries multiple frontier models in parallel         │
│ - Averages token probabilities or votes on answers     │
│ - Cross-verifies responses to reduce hallucinations    │
└─────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────┐
│ Stage 4: Confidence Scoring                             │
│ - Evaluates model confidence in response               │
│ - High confidence (>0.8): return response              │
│ - Low confidence (<0.8): escalate to premium model     │
└─────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────┐
│ Stage 5: Response Transformation                        │
│ - Sanitizes reasoning content (chain-of-thought)       │
│ - Applies AVELIN branding                              │
│ - Strips provider identifiers                          │
└─────────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────────┐
│ Stage 6: Continuous Evaluation (background)             │
│ - Benchmarks all providers hourly                      │
│ - Tracks: accuracy, latency, cost per task type        │
│ - Auto-updates selection weights based on performance    │
└─────────────────────────────────────────────────────────┘
    ↓
Response Returned (3.5s average total latency)
```

## Stage 1: Semantic Routing (50ms Latency)

AVELIN's semantic router classifies every incoming request before it reaches a model. This is the key differentiator from basic multi-provider proxies.

**How it works:**
1. Request text is embedded using a lightweight semantic transformer model
2. KNN classifier matches to task type based on training data
3. Optimal model family is selected

**Task classification examples:**

| Request | Classification | Route To |
|---------|---------------|----------|
| "Write Python code for..." | CODE | avelin-coding-fast |
| "Solve this calculus problem..." | REASONING | avelin-ultra |
| "Tell me a joke..." | SIMPLE_CHAT | avelin-fast |
| "Analyze this contract..." | DOCUMENT_ANALYSIS | avelin-pro |
| "Debug this error..." | CODE_DEBUG | avelin-coding-fast |

**Impact:**
- 68% of requests routed to cost-efficient models
- 23% quality improvement on specialized tasks
- 50ms classification latency (negligible vs 2-5s model response)

**Training data:** Built from millions of real production requests, continuously updated.

## Stage 2: Intelligent Model Selection

Once the task type is classified, AVELIN's router selects the optimal model tier within that family.

**Selection criteria:**
- **Cost:** Prefer cheaper models for simple tasks
- **Latency:** Consider time-to-first-token requirements
- **Quality:** Match model capability to task complexity
- **Availability:** Route away from degraded providers
- **Context length:** Ensure model can handle request size

**Model tier mapping:**

| Family | Ultra (Premium) | Pro (Balanced) | Fast (Efficient) |
|--------|----------------|----------------|------------------|
| Intelligence | Complex reasoning, strategy | Daily workflows, analysis | Simple chat, classification |
| Coding | Architecture, security review | Feature development | Bug fixes, refactoring |
| Agentic | Long-horizon planning | Standard tool use | High-volume agent swarms |

## Stage 3: Multi-Model Fusion (Optional)

For critical requests, AVELIN can query multiple frontier models in parallel and aggregate their responses.

**Aggregation strategies:**

1. **Weighted token probability averaging**
   - Combine probability distributions from multiple models
   - Higher weight to models with better benchmark scores
   - Produces "average" response that's often better than any single model

2. **Voting**
   - Each model generates independent response
   - Pick most common answer (majority vote)
   - Effective for factual questions

3. **Synthesis**
   - Use a meta-model to merge responses
   - Extract best parts from each
   - Most expensive but highest quality

**Impact:**
- 17% hallucination reduction through cross-verification
- Stronger synthesis on ambiguous tasks
- Higher confidence in critical responses

**When enabled:** Automatically for avelin-ultra on high-stakes requests, or manually via API parameter.

## Stage 4: Confidence-Based Cascading

AVELIN doesn't always use the most expensive model. It starts with cost-efficient models and escalates only when needed.

**The cascading flow:**

```
Request arrives
    ↓
avelin-fast tries (cost-efficient tier, $0.0005/request)
    ↓
Confidence score calculated (0.0 to 1.0)
    ↓
┌──────────────────────────────────────┐
│ If confidence > 0.8:                 │
│   Return response (cheap path)       │
│                                      │
│ If confidence < 0.8:                 │
│   Escalate to avelin-ultra           │
│   ($0.004/request)                   │
│   Return premium response            │
└──────────────────────────────────────┘
```

**Impact:**
- up to 85% cost savings on routine workloads
- Premium quality preserved for complex tasks
- Automatic optimization based on model uncertainty

**Escalation rate:** ~20% of requests escalate (80% handled by efficient models).

## Stage 5: Response Transformation

Before returning the response to the user, AVELIN applies several transformations:

1. **Reasoning sanitization:** Strips chain-of-thought that might reveal backend providers
2. **Branding:** Ensures consistent AVELIN identity
3. **Error handling:** Sanitizes error messages to prevent information leakage
4. **Format normalization:** Ensures consistent API response format

## Stage 6: Continuous Evaluation (Background)

AVELIN continuously benchmarks all providers and automatically updates selection weights.

**Evaluation process:**
1. Every hour, run benchmark suite on all providers
2. Track: accuracy, latency, cost, availability per task type
3. Update selection weights based on performance
4. Deploy updated weights (no restart required)

**Real-world results (last 90 days):**
- Routing weights updated 312 times
- 8% quality improvement over baseline
- 12% cost decrease as cheaper models improved
- Automatic failover when providers degraded

**Benchmark categories:**
- Coding: SWE-bench, HumanEval, MBPP
- Reasoning: MMLU, GPQA, GSM8K
- Agentic: TauBench, AgentBench
- Multilingual: Various language tasks

## Performance Characteristics

**Latency breakdown (median):**
- Prompt-vector classifier: 50ms
- Model selection: 10ms
- Model inference: 2,500ms (varies by provider)
- Response transformation: 20ms
- **Total: ~2.6s average**

**Streaming throughput:** 18.8 chunks/sec

**Time-to-first-token:** 1.6s median

**Reliability:**
- 99.97% uptime (multi-provider fallback)
- 0.3s average failover time
- 0 data loss incidents in 12 months

## Cost Optimization

**Average cost per request:** $0.003 (varies by model tier)

**Cost breakdown by specialist selection:**
- 68% requests → efficient models ($0.0002-$0.0005)
- 20% requests → balanced models ($0.001-$0.002)
- 12% requests → premium models ($0.002-$0.007)

**vs alternatives:**
- vs single-provider direct access: 42% cheaper
- vs always-premium selection: up to 85% cheaper
- vs manual model selection: 23% better quality on specialized tasks

## Comparison to Traditional Approaches

| Approach | How It Works | Limitations |
|----------|--------------|-------------|
| **Single provider** | All requests to one model | Expensive, single point of failure, vendor lock-in |
| **Manual selection** | User picks model per request | Requires expertise, suboptimal, high cognitive load |
| **Static rules** | "Code requests → coding model" | No adaptation, no optimization over time |
| **Basic proxy** | Round-robin or weighted traffic splitting | No task classification, no quality optimization |
| **AVELIN** | Prompt-vector classifier + fusion + difficulty-based cascading + continuous evaluation | **Optimal cost, quality, and reliability** |

## Technical Specifications

**Supported protocols:**
- OpenAI-compatible API (chat completions, embeddings)
- Function calling / tool use
- Streaming responses
- Batch processing

**Model families:**
- Intelligence: ultra, pro, fast
- Coding: coding, coding-plus, coding-architect
- Agentic: agentic-high, agentic, agentic-fast
- Utility: embeddings (bge-m3), transcription (whisper), image generation

**Infrastructure:**
- Kubernetes-based (k3s cluster)
- Multi-region deployment (Hetzner, OVHcloud)
- Horizontal pod autoscaling (3-10 replicas)
- Redis for distributed state
- PostgreSQL for persistence

**Security:**
- TLS encryption in transit
- API key authentication
- Request/response sanitization
- PII stripping (configurable)
- Zero data retention policy (configurable)

## Getting Started

AVELIN is OpenAI-compatible. Switch in 5 minutes:

```python
from openai import OpenAI

client = OpenAI(
    base_url="https://api.avelin.ai/v1",
    api_key="your-avelin-api-key"
)

response = client.chat.completions.create(
    model="avelin-pro",  # or ultra, fast, coding, etc.
    messages=[{"role": "user", "content": "Hello!"}]
)
```

No code changes needed if you're already using the OpenAI SDK. AVELIN handles all the specialist selection, fusion, and optimization automatically using its prompt-vector classifier with 90% accuracy.

## Related Documentation

- [Platform Overview](../platform-overview.md)
- [Model Catalog](../models/README.md)
- [API Reference](../api/index.md)
- [Competitive Advantages](../benefits/competitive-advantages.md)

---

URL: https://avelin.ai/docs/benefits

# Benefits and Advantages

This section explains how AVELIN creates measurable business value across three dimensions: strategic positioning, operational execution, and competitive differentiation.

AVELIN is an **AI Intelligence Platform** built on three pillars — **AI Lab**, **Cloud Token Factory**, and **Enterprise Platform** — and powered by **Cross-Model MoE** technology that fuses open models into 9 specialist families. This architecture delivers:

- **up to 85% lower cost** than frontier alternatives for intelligence workloads
- **up to 86% lower cost** than frontier alternatives for coding workloads
- **up to 87% lower cost** than frontier alternatives for agentic workloads

AVELIN models are continuously evaluated against independent benchmarks; results are published at eval.avelin.ai.

## How to Use This Section

- Use this folder to align leadership, operations, and go-to-market teams on one value narrative
- Start with strategic value if you are defining direction or investment priorities
- Start with operational value if you are planning rollout, adoption, or KPI tracking
- Use competitive value for positioning, differentiation, and executive communication
- Use benchmark results for data-backed claims, sales decks, and technical evaluations

## Benefit Pages

- [`platform-strategic-benefits.md`](platform-strategic-benefits.md) — Strategic value from the 3-pillar architecture
- [`operational-benefits.md`](operational-benefits.md) — Day-to-day execution improvements with real numbers
- [`competitive-advantages.md`](competitive-advantages.md) — Market positioning with cost and quality comparisons
- [`benchmark-results.md`](benchmark-results.md) — How AVELIN models are evaluated

## Benefit Architecture at a Glance

| Benefit Layer | Core Question | Primary Audience | Typical Outcomes |
| --- | --- | --- | --- |
| Strategic | Why does this matter at company level? | Executive leadership, strategy, governance | up to 87% cost reduction, sovereign AI ownership, vendor-agnostic orchestration |
| Operational | How does this improve daily execution? | Business units, operations, delivery teams | Faster workflows, independently verified quality, lower per-request cost |
| Competitive | Why do we win with this approach? | Leadership, product, sales, partnerships | Top 1–4% benchmark scores at fraction of competitor pricing |
| Benchmark | What does the data say? | Engineering, procurement, technical evaluators | Independent evaluation results, cost-per-token analysis, methodology transparency |

## Recommended Reading Order

1. [`benchmark-results.md`](benchmark-results.md) for how AVELIN models are evaluated — detailed results at eval.avelin.ai
2. [`platform-strategic-benefits.md`](platform-strategic-benefits.md) for leadership outcomes and long-range value
3. [`operational-benefits.md`](operational-benefits.md) for team-level execution gains and measurable efficiency
4. [`competitive-advantages.md`](competitive-advantages.md) for market positioning and differentiation

## Key Numbers at a Glance

| Metric | Value |
| --- | --- |
| Intelligence family cost advantage | **up to 85% cheaper** than frontier alternatives |
| Coding family cost advantage | **up to 86% cheaper** than frontier alternatives |
| Agentic family cost advantage | **up to 87% cheaper** than frontier alternatives |
| Specialist model families | **9** (via Cross-Model MoE) |
| Platform pillars | **3** (AI Lab, Cloud Token Factory, Enterprise Platform) |
| Evaluation | Continuous, independent — results at eval.avelin.ai |

## Suggested Review Rhythm

- Revisit strategic benefits quarterly
- Revisit operational benefits monthly with KPI updates
- Revisit competitive advantages each planning cycle or market shift
- Revisit evaluation results regularly at eval.avelin.ai

---

URL: https://avelin.ai/docs/operations

# Operations and Governance

This section defines how to run AVELIN's **AI Intelligence Platform** reliably, securely, and at enterprise scale.

Operations span all **3 pillars** (AI Lab, Cloud Token Factory, Enterprise Platform) and leverage **Cross-Model MoE** to deliver:

- **Intelligence up to 85% cheaper** than premium alternatives
- **Coding up to 86% cheaper** through optimized cascading
- **Agentic up to 87% cheaper** via intelligent tier selection

It is built for teams that must keep productivity high while maintaining strong governance, data privacy, and service continuity.

## Operations Objectives

- Keep AVELIN-GPT, AVELIN-API, AVELIN-MCP, and Y-RAY stable for critical business workflows
- Enforce trust, access, and policy controls across all products
- Detect and resolve incidents quickly with clear ownership
- Track KPI trends that show reliability, adoption, quality, and governance maturity
- Improve operations continuously without slowing business execution

## Operations Pages

- [`security-reliability-governance.md`](security-reliability-governance.md): governance model, controls, cadence, accountability
- [`deployment-and-administration.md`](deployment-and-administration.md): deployment principles, admin runbooks, continuity readiness
- [`troubleshooting-and-kpis.md`](troubleshooting-and-kpis.md): incident triage model, issue matrix, KPI framework

## Primary Audience

- Platform administrators
- Security and governance leaders
- Operations managers and service owners
- Business owners responsible for AI-enabled workflow outcomes

## Operating Rhythm

| Cadence | Primary Focus | Typical Outputs |
| --- | --- | --- |
| Daily | Service stability and urgent issue handling | Incident actions, quick risk controls, recovery confirmations |
| Weekly | Reliability and governance execution | Health review, exception review, owner follow-ups |
| Monthly | KPI trends and optimization priorities | Performance insights, governance updates, improvement plan |
| Quarterly | Maturity and strategic alignment | Risk posture review, investment priorities, roadmap decisions |

## Recommended Reading Path1. Start with [`security-reliability-governance.md`](security-reliability-governance.md) to define control expectations.
2. Continue with [`deployment-and-administration.md`](deployment-and-administration.md) to operationalize ownership and routines.
3. Use [`troubleshooting-and-kpis.md`](troubleshooting-and-kpis.md) to run incidents and measure outcomes consistently.

---

URL: https://avelin.ai/docs/mcp

# AVELIN-MCP Services Documentation

This section provides complete business-facing documentation for AVELIN-MCP service actions and model-achievable workflows.

AVELIN-MCP is part of the **Enterprise Platform pillar** and executes actions through **Cross-Model MoE** intelligence — delivering agentic workflows **up to 87% cheaper** than premium alternatives by cascading through optimal tiers.

Coverage is based on implemented MCP actions for:

- Google Calendar
- Gmail
- Google Contacts
- Google Tasks

## What You Will Find Here

- Full per-service action catalogs with exact tool names
- Business description of what each action does
- Practical scenarios the model can execute with each service
- Cross-service automation scenarios combining Calendar, Gmail, Contacts, and Tasks
- Quick action reference for fast lookup

## Documentation Map

- Google Calendar actions and scenarios: [`google-calendar.md`](google-calendar.md)
- Gmail actions and scenarios: [`gmail.md`](gmail.md)
- Google Contacts actions and scenarios: [`google-contacts.md`](google-contacts.md)
- Google Tasks actions and scenarios: [`google-tasks.md`](google-tasks.md)
- Cross-service workflows: [`cross-service-scenarios.md`](cross-service-scenarios.md)
- Full quick reference: [`action-reference.md`](action-reference.md)

## Action Coverage Summary

| Service | Action Count |
| --- | --- |
| Google Calendar | 13 |
| Gmail | 32 |
| Google Contacts | 9 |
| Google Tasks | 15 |
| **Total** | **69** |

## How to Use This Section

1. Start with the service page that matches your immediate business workflow.
2. Use the action descriptions to understand exactly what can be automated.
3. Use per-service scenarios to design role-specific playbooks.
4. Use cross-service scenarios to build end-to-end execution flows.

## Related Platform Pages

- AVELIN-MCP overview: [`../systems/mcp-integration-platform.md`](../systems/mcp-integration-platform.md)
- Use-case index: [`../use-cases/index.md`](../use-cases/index.md)

---

URL: https://avelin.ai/docs/glossary

# Business Glossary

This glossary defines business and operating terms used across AVELIN documentation.

## A

**A.V.E.L.I.N.**  
Artificial Virtual Entity Learning Interaction Natively. The sovereign AI platform for business.

**Action Superpowers**  
The ability to move from analysis to execution through connected business actions, not only chat responses.

**Adaptive Threat Defense**  
Perimeter behavior that detects suspicious traffic patterns and reduces harmful traffic impact before core services are affected.

**Anonymization Layer**  
A trust control that removes sensitive identity patterns before external processing when policy requires it.

**AI Intelligence Platform**  
AVELIN's positioning as a sovereign AI operating layer for business, organized around 3 pillars and Cross-Model MoE technology.

**AI Lab**  
One of AVELIN's 3 pillars — advanced reasoning, research, and decision-grade intelligence capabilities.

**Audit Readiness**  
The ability to produce complete, timely governance and activity evidence for internal and external review.

**Avatar Marketplace**  
A catalog of specialized AI personas and domain-focused assistants that can be activated for specific tasks.

**Evaluation**  
AVELIN models are continuously evaluated against independent benchmarks; results are published on the dedicated evaluation portal, eval.avelin.ai.

**AVELIN-API**  
The orchestration and governance layer that fuses and routes model intelligence.

**AVELIN-GPT**  
The conversational workspace where users analyze, decide, automate, and execute daily workflows.

**AVELIN-MCP**  
The integration platform that extends AVELIN with high-performance execution capabilities.

## B

**Business Continuity**  
The capability to keep critical workflows available during incidents, traffic stress, or platform changes.

**Business Workflow Criticality**  
The business importance level of a workflow, used to prioritize incident response and governance controls.

## C

**Classic Mode**  
Standard question-and-answer mode for routine interactions.

**Cloud Token Factory**  
One of AVELIN's 3 pillars — cost-optimized, high-throughput model orchestration and token governance. Delivers Intelligence up to 85%, Coding up to 86%, and Agentic up to 87% cheaper than premium alternatives.

**Compliance Evidence**  
Documented records that demonstrate policy adherence, control execution, and governance quality.

**Corporate Brain**  
Enterprise knowledge intelligence built from internal documents, records, and institutional memory.

**Cross-Model MoE**  
AVELIN's core technology that combines multiple intelligence models into unified outputs using a trained meta-model, not a router. A prompt-vector classifier with 90% accuracy assigns each request to the optimal tier through difficulty-based cascading for cost and quality balance.

**Cross-System Trust Enforcement**  
Consistent trust and policy controls applied across AVELIN-GPT, AVELIN-API, AVELIN-MCP, and Y-RAY.

## D

**Data Residency**  
The defined location or boundary where sensitive data is processed and retained.

**Data Sovereignty**  
The principle that organizations keep control of where data is processed, stored, and governed.

**Debate Mode**  
A multi-perspective reasoning mode where contrasting intelligence paths improve balance and rigor.

**Decision-Grade Output**  
An output format suitable for high-impact decisions, with evidence quality and policy alignment.

**Distributed Intelligence**  
An operating model where multiple intelligence capabilities work together as one governed platform.

## E

**Enterprise Platform**  
One of AVELIN's 3 pillars — integration, execution, trust, and operations for business workflows.

**Enterprise Trust Layer**  
The combined policy, access, retention, and quality-control capabilities that keep AI usage safe and auditable.

**Evidence-First Intelligence**  
A decision approach that prioritizes verifiable signals and traceable support over unsupported output.

**Execution Workflow**  
A structured path that connects analysis, approval, and action completion for a business task.

## G

**Governance Cadence**  
The scheduled rhythm for control reviews, KPI reviews, risk checks, and policy decisions.

**Governed Channel**  
A platform path where trust, policy, and access controls are actively enforced and monitored.

**Guardrails**  
Rules that keep behavior, outputs, and actions aligned with business policy and risk standards.

## H

**Hybrid Deployment**  
A deployment model that combines strict internal zones for sensitive workloads with broader infrastructure for lower-risk workloads.

## I

**Incident Severity**  
A business-impact classification that determines response urgency, escalation level, and owner involvement.

**Integration Governance**  
Control of connector permissions, ownership, and safe action execution across integrated systems.

**Intelligence Governance**  
Control of routing, cost, output quality, and policy alignment for model-driven workflows.

## K

**KPI (Key Performance Indicator)**  
A measurable signal used to evaluate reliability, productivity, quality, governance, and business outcomes.

## L

**Least Privilege**  
An access model where users and services receive only the permissions needed for their responsibilities.

**Local Server Deployment**  
A sovereignty-first deployment model where AVELIN runs inside organization-controlled infrastructure.

## M

**MCP (Model Context Protocol)**  
A standard way for AI systems to access external tools and business services in controlled workflows.

**Memory Mode**  
An interaction mode that uses persistent context to improve continuity and personalization over time.

**Model Alias**  
A business-friendly model name used inside AVELIN to simplify routing and user experience.

**Multi-Depth Analysis**  
A structured approach to research depth, from rapid scan to deep diligence depending on decision importance.

## P

**Perimeter Layer**  
The secure entry boundary that applies transport protection, traffic controls, and threat defense.

**Policy Exception**  
A temporary, approved deviation from a standard control, with owner, reason, and expiry.

**Private Cloud Deployment**  
A controlled cloud model that keeps governance boundaries while supporting enterprise scale.

## R

**Research-to-Execution Continuum**  
The end-to-end path from intelligence gathering to business action completion in one governed flow.

**Retention Policy**  
The rule set that defines how long data, logs, and evidence are kept before deletion.

**Route Policy**  
Rules that determine how requests are directed across intelligence paths based on quality, cost, and risk.

## S

**Smart Routing Economy**  
A routing strategy that sends each task to the most appropriate intelligence path for quality and cost efficiency.

**Sovereign AI Platform**  
An AI operating model where the organization controls infrastructure, governance, and critical data boundaries. AVELIN is positioned as an AI Intelligence Platform with 3 pillars and Cross-Model MoE technology.

**Sovereign Deployment**  
A deployment posture where control over data boundaries, policy enforcement, and trust operations remains with the organization.

## T

**Token Counting**  
Measurement of usage volume for cost tracking, governance, and optimization.

**Traceability**  
The ability to follow who performed what action, when, and under which policy context.

**Trust Controls**  
Security, policy, privacy, and quality controls that reduce risk in sensitive AI workflows.

## V

**Vendor Agnostic Intelligence**  
The ability to use and switch between multiple model families without being locked to one provider path.

## Y

**Y-RAY**  
The information search and analysis product that combines internal corporate intelligence with external research intelligence.

**Y-RAY Alpha Scan**  
Rapid intelligence mode for quick pulse updates and immediate signal detection.

**Y-RAY Deep Diligence**  
Investigation mode for claim-by-claim validation, contradiction detection, and decision-grade synthesis.

**Y-RAY Structural Breakdown**  
Analytical mode for mapping relationships, risks, dependencies, and strategic structure.

## Z

**Zero-Retention Environment**  
A controlled processing mode where persistent history can be minimized or disabled for sensitive workflows.

---

URL: https://avelin.ai/docs/pricing

# Pricing

**Pay per token, only for what you use.** No subscriptions, no minimums, no hidden fees.

All prices are in USD per 1 million tokens.

## Intelligence

General reasoning, writing, and vision · 256K context

| Model | Input | Output | Cache write | Cache read |
|---|---|---|---|---|
| `avelin-fast` | $0.30 | $1.20 | $0.38 | $0.03 |
| `avelin-pro` | $1.40 | $4.40 | $1.75 | $0.26 |
| `avelin-ultra` | $2.50 | $12.50 | $3.13 | $0.52 |

## Coding

Software engineering · 1M context — an entire repository in one request

| Model | Input | Output | Cache write | Cache read |
|---|---|---|---|---|
| `avelin-coding-fast` | $0.30 | $1.20 | $0.38 | $0.03 |
| `avelin-coding-pro` | $1.40 | $4.40 | $1.75 | $0.26 |
| `avelin-coding-ultra` | $2.50 | $12.50 | $3.13 | $0.52 |

## Agentic

Tool calling and autonomous agents · 256K context

| Model | Input | Output | Cache write | Cache read |
|---|---|---|---|---|
| `avelin-agentic-fast` | $0.285 | $1.14 | $0.36 | $0.0285 |
| `avelin-agentic-pro` | $1.26 | $3.96 | $1.58 | $0.23 |
| `avelin-agentic-ultra` | $2.38 | $11.88 | $2.98 | $0.49 |

**fast** for volume · **pro** for everyday production · **ultra** for the hardest problems. Not sure? Start with `avelin-pro`. Capabilities and selection guide: [Model Catalog](models/README.md).

**Prompt caching:** repeated prompt prefixes (system prompts, documents, history) are billed once at the *cache write* rate, then at the *cache read* rate on every reuse — roughly 5× below input on the pro and ultra tiers, and 10× below on all three fast tiers.

**Legacy names:** `avelin-coding`, `avelin-coding-plus`, `avelin-coding-architect`, `avelin-agentic`, and `avelin-agentic-high` still work and are billed as their [renamed models](models/README.md#legacy-model-names).

## Audio, images, and embeddings

| Model | What it does | Price |
|---|---|---|
| `bge-m3` | Embeddings for search and RAG | $0.02 per 1M tokens |
| `whisper-large-v3` | Speech-to-text, highest accuracy | $0.006 per minute |
| `whisper-large-v3-turbo` | Speech-to-text, fastest | $0.006 per minute |
| `avelin-stt` | Speech-to-text, streaming-friendly | $0.006 per minute |
| `tts-1` | Text-to-speech, low latency | $0.12 per 1K characters |
| `tts-1-hd` | Text-to-speech, highest quality | $0.24 per 1K characters |
| `avelin-imagegen` | Image generation, standard | $0.03 per image |
| `avelin-imagegen-pro` | Image generation, high quality | $0.075 per image |

## Web Intelligence

Firecrawl-compatible [web access](api/web-scraping.md), billed per request. `crawl` and `extract` are asynchronous jobs, charged once at kickoff.

| Endpoint | Operation | Price per request |
|---|---|---|
| `POST /v2/scrape` | Scrape a single URL | $0.001 |
| `POST /v2/map` | Map a site's links | $0.001 |
| `POST /v2/search` | Web search | $0.005 |
| `POST /v2/crawl` | Start a crawl job | $0.010 |
| `POST /v2/extract` | Start an extract job | $0.005 |

## Estimating costs

1,000 tokens ≈ 750 words. Summarizing a 2,000-word document with `avelin-pro` — roughly 2,700 input and 400 output tokens — costs about **$0.006**, or $5.60 per 1,000 documents.

## Billing

- Metered per token, invoiced monthly with a full breakdown
- Live usage dashboard with budget alerts; every API key can carry its own budget and rate limits
- One price per model — failover to a backup never changes your bill
- No commitments, cancel anytime

## Enterprise and sovereign deployment

Run AVELIN on your own GPUs — on-premises, private cloud, or air-gapped. Unlimited usage under fixed annual licensing, complete data sovereignty, compliance-ready (HIPAA, GDPR, SOX, FedRAMP), continuous model updates with enterprise SLA.

Pricing is tailored to your deployment — **[request a POC](/poc#request)**. Trials available on request.

## Get started

1. Request your API key — **sales@avelin.ai**
2. Point your OpenAI- or Anthropic-compatible client at `https://api.avelin.ai` — a [2-line change](api/quickstart.md)
3. Send your first request to `avelin-pro`

Volume pricing and sovereign deployment: **[Request a POC](/poc#request)**

---

URL: https://avelin.ai/docs/models/avelin-fast

# avelin-fast

## Model Positioning

`avelin-fast` is AVELIN's low-latency, high-throughput tier — optimized for speed and cost efficiency in routine, high-volume work. Built with AVELIN's Cross-Model MoE technology in our AI Lab, it fuses multiple model architectures into a single trained system optimized for speed behind a single endpoint.

**Context window:** 256K tokens input, 65K tokens output

**Pricing:** $0.30 input / $1.20 output per 1M tokens (cache read $0.03, cache write $0.38) — see [Pricing](../pricing.md)

**Vision support:** When your request contains images, avelin-fast automatically routes to a vision-capable model behind the scenes. You send to `avelin-fast` — we handle the rest. No special model selection needed.

Choose this tier when you need quick, dependable answers at scale.

## Cross-Model MoE

Our Cross-Model MoE uses a prompt-vector classifier trained on 1 year of production data to predict the optimal specialist with 90% accuracy. It applies difficulty-based cascading (simple tasks → fast models, complex tasks → multi-expert fusion) and cost-aware routing (same quality, cheapest path). This is not a router — routers pick one model via if/else rules. We cascade across trained experts and synthesize output quality above any single expert.
Built with AVELIN's Cross-Model MoE technology in our AI Lab. Cross-Model MoE is a trained meta-model — not a router — that fuses knowledge from multiple model architectures into a single unified model. You call `avelin-fast` and get fast output based on task type and latency requirements. For image requests, vision capabilities are engaged automatically. No model selection needed.

**Self-improving platform:** As new frontier models are released, AVELIN integrates them within 48 hours. Your `avelin-fast` endpoint gets faster and smarter over time — automatically, with zero code changes.

**Enterprise licensing:** AVELIN models are available for sovereign deployment on customer GPUs. [Request a POC](/poc#request) for enterprise licensing — run AVELIN on your own infrastructure with full data sovereignty.

## Capabilities at a Glance

| Capability | Detail |
| --- | --- |
| Context window | **256K tokens** |
| Max output | **65K tokens** |
| Reasoning | off by default (optimized for speed) |
| Tools / function calling | yes |
| Streaming | yes |
| Vision | ✅ yes (auto-routed for image requests) |

## Best-Fit Workloads

- High-volume drafting and rewriting
- Classification, extraction, and tagging
- Chat and assistant interactions where latency matters
- Batch processing of routine requests

## When to Choose avelin-fast

- The task is routine and well-defined
- You are processing many requests and care about cost and speed
- Deep multi-step reasoning is not required

For ambiguous or high-stakes work, escalate to `avelin-pro` or `avelin-ultra`.

## Example Business Requests

- "Summarize these 50 support tickets into one line each."
- "Rewrite this paragraph in a friendlier tone."
- "Classify these messages as billing, technical, or sales."

## Related

- [Model Catalog](README.md)
- [API Reference](../api/index.md)

---

URL: https://avelin.ai/docs/models/avelin-pro

# avelin-pro

## Model Positioning

`avelin-pro` is AVELIN's high-efficiency reasoning tier — delivering strong intelligence with faster turnaround and lower cost. Built with AVELIN's Cross-Model MoE technology in our AI Lab, it fuses multiple model architectures into a single trained system optimized for production economics.

**Context window:** 256K tokens input, 65K tokens output

**Pricing:** $1.40 input / $4.40 output per 1M tokens (cache read $0.26, cache write $1.75) — see [Pricing](../pricing.md)

Choose this tier when throughput, responsiveness, and routine workflow acceleration are the primary goals.

## Cross-Model MoE

Our Cross-Model MoE uses a prompt-vector classifier trained on 1 year of production data to predict the optimal specialist with 90% accuracy. It applies difficulty-based cascading (simple tasks → fast models, complex tasks → multi-expert fusion) and cost-aware routing (same quality, cheapest path). This is not a router — routers pick one model via if/else rules. We cascade across trained experts and synthesize output quality above any single expert.
Built with AVELIN's Cross-Model MoE technology in our AI Lab. Cross-Model MoE is a trained meta-model — not a router — that fuses knowledge from multiple model architectures into a single unified model. You call `avelin-pro` and get consistently strong output at production-friendly pricing without vendor lock-in.

**Self-improving platform:** As new frontier models are released, AVELIN integrates them within 48 hours. Your `avelin-pro` endpoint gets smarter over time — automatically, with zero code changes.

**Enterprise licensing:** AVELIN models are available for sovereign deployment on customer GPUs. [Request a POC](/poc#request) for enterprise licensing — run AVELIN on your own infrastructure with full data sovereignty.

## Capabilities at a Glance

| Capability | Detail |
| --- | --- |
| Context window | **256K tokens** |
| Max output | **65K tokens** |
| Reasoning | high effort |
| Tools / function calling | yes |
| Streaming | yes |
| Vision | yes |

## Best-Fit Workloads

- Daily communication drafting
- Meeting preparation and recap generation
- Task planning and follow-up sequencing
- Internal knowledge Q&A
- Repetitive operational workflows at scale

## Functional Strengths

### Fast Response Performance

Delivers quick turnaround for iterative workflows and high request volume.

### Strong Day-to-Day Utility

Excellent for practical business tasks that need speed and consistency.

### Workflow Throughput

Supports large volumes of routine prompts without sacrificing usability.

### Cost-Efficient Execution

Well suited for Smart Routing Economy policies in AVELIN-API. `avelin-pro` is the go-to tier for high-volume production workloads.

## Business Benefits

- Higher team productivity in daily operations
- Faster communication and coordination cycles
- Lower cost per routine workflow
- Better scalability for organization-wide adoption

## When to Choose avelin-pro

Use `avelin-pro` when these conditions apply:

- The task is routine, repeatable, and operational in nature
- Speed and turnaround are critical
- Output is important but does not require deep strategic reasoning
- You are processing large volumes of similar requests

## Prompting Guidance

- Keep instructions clear and action-oriented
- Specify output format to reduce revision cycles
- For complex ambiguity, escalate to `avelin-ultra`

## Example Business Requests

- "Draft a customer follow-up email from these meeting notes."
- "Summarize this conversation into five action items and owners."
- "Generate tomorrow's priority plan from this task list."

## Success Metrics

- Faster average completion time for routine tasks
- Increased daily workflow throughput per user
- Lower spend for operational prompt volume

---

URL: https://avelin.ai/docs/models/avelin-ultra

# avelin-ultra

## Model Positioning

`avelin-ultra` is AVELIN's flagship reasoning tier — the highest-intelligence model in the Intelligence family. Built with AVELIN's Cross-Model MoE technology in our AI Lab, it fuses multiple model architectures into a single trained system for high-complexity business work.

**Context window:** 256K tokens input, 65K tokens output

**Pricing:** $2.50 input / $12.50 output per 1M tokens (cache read $0.52, cache write $3.13) — see [Pricing](../pricing.md)

Choose this tier when decision quality, rigor, and structured thinking are more important than raw speed.

## Cross-Model MoE

Our Cross-Model MoE uses a prompt-vector classifier trained on 1 year of production data to predict the optimal specialist with 90% accuracy. It applies difficulty-based cascading (simple tasks → fast models, complex tasks → multi-expert fusion) and cost-aware routing (same quality, cheapest path). This is not a router — routers pick one model via if/else rules. We cascade across trained experts and synthesize output quality above any single expert.
Built with AVELIN's Cross-Model MoE technology in our AI Lab. Cross-Model MoE is a trained meta-model — not a router — that fuses knowledge from multiple model architectures into a single unified model. You call `avelin-ultra` and get consistently top-tier output without model selection or vendor lock-in.

**Self-improving platform:** As new frontier models are released, AVELIN integrates them within 48 hours. Your `avelin-ultra` endpoint gets smarter over time — automatically, with zero code changes.

**Enterprise licensing:** AVELIN models are available for sovereign deployment on customer GPUs. [Request a POC](/poc#request) for enterprise licensing — run AVELIN on your own infrastructure with full data sovereignty.

## Capabilities at a Glance

| Capability | Detail |
| --- | --- |
| Context window | **256K tokens** |
| Max output | **65K tokens** |
| Reasoning | deepest (high effort) |
| Tools / function calling | yes |
| Streaming | yes |
| Vision | yes |

## Best-Fit Workloads

- Strategic planning and scenario analysis
- Executive decision support
- Risk evaluation and trade-off analysis
- Complex synthesis from multiple information streams
- High-stakes communications that require precision

## Functional Strengths

### Deep Multi-Step Reasoning

Handles layered problems with multiple dependencies and constraints.

### Strong Synthesis Quality

Combines large context and mixed input types into coherent decision-ready outputs.

### Nuanced Decision Framing

Produces structured options, consequences, and recommendation logic for leadership use.

### High-Reliability Output Style

Optimized for quality and completeness in critical workflows.

## Business Benefits

- Better quality in strategic and high-impact decisions
- Reduced rework on complex deliverables
- Higher confidence in leadership-facing outputs
- Stronger consistency for policy-sensitive or risk-sensitive tasks

## When to Choose avelin-ultra

Use `avelin-ultra` when at least one of these is true:

- The decision has meaningful financial, legal, or reputational impact
- The task has ambiguity and no obvious single-path answer
- The output needs board, executive, or client-grade quality
- You need explicit option comparison and recommendation structure

## Prompting Guidance

- Ask for structured outputs (options, risks, recommendation, next actions)
- Provide constraints and success criteria up front
- Request confidence framing when decisions involve uncertainty

## Example Business Requests

- "Build a 90-day strategic options memo with risk matrix and recommendation."
- "Compare three operating models and propose the best path with trade-offs."
- "Create an executive brief with decision options and implementation sequence."

## Success Metrics

- Reduced revision cycles on strategic documents
- Higher stakeholder acceptance on first draft
- Faster decision closure for complex initiatives

---

URL: https://avelin.ai/docs/models/avelin-coding-fast

# avelin-coding-fast

> **Formerly `avelin-coding`.** The legacy name keeps working as a server-side alias — no client change required.

## Model Positioning

`avelin-coding-fast` is AVELIN's standard software-engineering tier — the balanced entry point in the Coding family for everyday agentic development workflows. Built with AVELIN's Cross-Model MoE technology in our AI Lab, it fuses multiple model architectures into a single trained system optimized for cost and capability.

**Context window:** 1M tokens input, 65K tokens output

**Pricing:** $0.30 input / $1.20 output per 1M tokens (cache read $0.03, cache write $0.38) — see [Pricing](../pricing.md)

Balanced engineering tier for day-to-day coding at the strongest cost-per-capability ratio in the family.

## Cross-Model MoE

Our Cross-Model MoE uses a prompt-vector classifier trained on 1 year of production data to predict the optimal specialist with 90% accuracy. It applies difficulty-based cascading (simple tasks → fast models, complex tasks → multi-expert fusion) and cost-aware routing (same quality, cheapest path). This is not a router — routers pick one model via if/else rules. We cascade across trained experts and synthesize output quality above any single expert.
Built with AVELIN's Cross-Model MoE technology in our AI Lab. Cross-Model MoE is a trained meta-model — not a router — that fuses knowledge from multiple model architectures into a single unified model. You call `avelin-coding-fast` and get consistently strong code output based on language, complexity, and context requirements — without vendor lock-in.

**Self-improving platform:** As new frontier models are released, AVELIN integrates them within 48 hours. Your `avelin-coding-fast` endpoint gets more capable over time — automatically, with zero code changes.

**Enterprise licensing:** AVELIN models are available for sovereign deployment on customer GPUs. [Request a POC](/poc#request) for enterprise licensing — run AVELIN on your own infrastructure with full data sovereignty.

**Why 1M context matters:** Load entire repositories into context. Understand cross-file dependencies, trace execution paths across modules, and refactor with complete codebase awareness. This isn't just a large context window — it's whole-repository intelligence.

## Capabilities at a Glance

| Capability | Detail |
| --- | --- |
| Context window | **1M tokens** |
| Max output | **65K tokens** |
| Reasoning | off by default (enable per request) |
| Tools / function calling | yes |
| Streaming | yes |

## Best-Fit Workloads

- Writing and reviewing application code
- Refactoring and bug fixing
- Explaining unfamiliar code and writing tests
- Working across large files and multi-file context

## When to Choose avelin-coding-fast

- Day-to-day development tasks
- You want strong code quality without the heaviest tier
- The repository or context is large

For harder design work, use `avelin-coding-pro` or `avelin-coding-ultra`.

## Example Business Requests

- "Add input validation to this endpoint and write tests."
- "Find and fix the bug causing this stack trace."
- "Explain what this module does and suggest improvements."

## Related

- [`avelin-coding-pro.md`](avelin-coding-pro.md)
- [`avelin-coding-ultra.md`](avelin-coding-ultra.md)
- [Model Catalog](README.md)
- [API Reference](../api/index.md)

---

URL: https://avelin.ai/docs/models/avelin-coding-pro

# avelin-coding-pro

> **Formerly `avelin-coding-plus`.** The legacy name keeps working as a server-side alias — no client change required.

## Model Positioning

`avelin-coding-pro` is AVELIN's high-throughput coding tier — delivering enhanced reasoning for complex engineering tasks at production speed. Built with AVELIN's Cross-Model MoE technology in our AI Lab, it fuses multiple model architectures into a single trained system that maximizes coding quality and workload capacity simultaneously.

**Context window:** 1M tokens input, 65K tokens output

**Pricing:** $1.40 input / $4.40 output per 1M tokens (cache read $0.26, cache write $1.75) — see [Pricing](../pricing.md)

Maximized throughput for high-volume agentic loops, large multi-file refactors, and latency-sensitive CI/code-review pipelines where waiting on tokens kills developer experience.

## Cross-Model MoE

Our Cross-Model MoE uses a prompt-vector classifier trained on 1 year of production data to predict the optimal specialist with 90% accuracy. It applies difficulty-based cascading (simple tasks → fast models, complex tasks → multi-expert fusion) and cost-aware routing (same quality, cheapest path). This is not a router — routers pick one model via if/else rules. We cascade across trained experts and synthesize output quality above any single expert.
Built with AVELIN's Cross-Model MoE technology in our AI Lab. Cross-Model MoE is a trained meta-model — not a router — that fuses knowledge from multiple model architectures into a single unified model. You call `avelin-coding-pro` and get high-quality code at production speed based on task complexity, language, and throughput requirements — without vendor lock-in.

**Self-improving platform:** As new frontier models are released, AVELIN integrates them within 48 hours. Your `avelin-coding-pro` endpoint gets more capable over time — automatically, with zero code changes.

**Enterprise licensing:** AVELIN models are available for sovereign deployment on customer GPUs. [Request a POC](/poc#request) for enterprise licensing — run AVELIN on your own infrastructure with full data sovereignty.

**The 1M context advantage:** Process entire repositories without chunking. When you're refactoring across 50 files or analyzing a complex codebase, you need to see everything at once. This is whole-repository intelligence at production speed.

## Capabilities at a Glance

| Capability | Detail |
| --- | --- |
| Context window | **1M tokens** |
| Max output | **65K tokens** |
| Reasoning | deep step-by-step (thinking) |
| Tools / function calling | yes |
| Streaming | yes |

## Best-Fit Workloads

- Non-trivial feature implementation
- Multi-file refactoring with cross-cutting impact
- Tricky debugging that needs careful reasoning
- Code that must meet higher correctness and quality bars

## When to Choose avelin-coding-pro

- `avelin-coding-fast` is not producing the depth you need
- The change spans multiple files or has subtle edge cases
- Correctness matters more than raw speed

For system-level architecture and design, use `avelin-coding-ultra`.

## Example Business Requests

- "Implement this feature across the API and the client, with tests."
- "Track down this intermittent race condition and fix it safely."
- "Refactor this service to a cleaner pattern without breaking callers."

## Related

- [`avelin-coding-fast.md`](avelin-coding-fast.md)
- [`avelin-coding-ultra.md`](avelin-coding-ultra.md)
- [Model Catalog](README.md)
- [API Reference](../api/index.md)

---

URL: https://avelin.ai/docs/models/avelin-coding-ultra

# avelin-coding-ultra

> **Formerly `avelin-coding-architect`.** The legacy name keeps working as a server-side alias — no client change required.

## Model Positioning

`avelin-coding-ultra` is AVELIN's deep-thinking architecture tier — the most capable model in the Coding family for system-level design decisions. Built with AVELIN's Cross-Model MoE technology in our AI Lab, it fuses multiple model architectures into a single trained system for whole-repository analysis and architectural planning.

**Context window:** 1M tokens input, 65K tokens output

**Pricing:** $2.50 input / $12.50 output per 1M tokens (cache read $0.52, cache write $3.13) — see [Pricing](../pricing.md)

The strongest pick for whole-repo reasoning, system-design decisions, and long-horizon refactors where one wrong call costs hours.

## Cross-Model MoE

Our Cross-Model MoE uses a prompt-vector classifier trained on 1 year of production data to predict the optimal specialist with 90% accuracy. It applies difficulty-based cascading (simple tasks → fast models, complex tasks → multi-expert fusion) and cost-aware routing (same quality, cheapest path). This is not a router — routers pick one model via if/else rules. We cascade across trained experts and synthesize output quality above any single expert.
Built with AVELIN's Cross-Model MoE technology in our AI Lab. Cross-Model MoE is a trained meta-model — not a router — that fuses knowledge from multiple model architectures into a single unified model. You call `avelin-coding-ultra` and get maximized architectural insight and design quality — system-level intelligence without vendor lock-in.

**Self-improving platform:** As new frontier models are released, AVELIN integrates them within 48 hours. Your `avelin-coding-ultra` endpoint gets more capable over time — automatically, with zero code changes.

**Enterprise licensing:** AVELIN models are available for sovereign deployment on customer GPUs. [Request a POC](/poc#request) for enterprise licensing — run AVELIN on your own infrastructure with full data sovereignty.

**Why architects need 1M context:** System design requires understanding the entire codebase — dependencies, patterns, interfaces, and edge cases across hundreds of files. With 1M context, you can load your complete repository and reason about architecture holistically, not piecemeal. This is the difference between "let me check those 5 files" and "I understand your entire system."

## Capabilities at a Glance

| Capability | Detail |
| --- | --- |
| Context window | **1M tokens** |
| Max output | **65K tokens** |
| Reasoning | deepest step-by-step (thinking) |
| Tools / function calling | yes |
| Streaming | yes |

## Best-Fit Workloads

- System and API design decisions
- Architecture reviews and trade-off analysis
- Large-scale refactoring and migration planning
- High-stakes engineering work where design quality drives outcomes

## When to Choose avelin-coding-ultra

- The problem is about design and structure, not just implementation
- You need explicit trade-off analysis and a recommended approach
- The decision has long-lived architectural impact

For implementation once the design is set, `avelin-coding-pro` or
`avelin-coding-fast` are more cost-efficient.

## Example Business Requests

- "Design a scalable event-driven architecture for this workload."
- "Review this system design and identify risks and alternatives."
- "Plan a migration from monolith to services with a phased roadmap."

## Related

- [`avelin-coding-pro.md`](avelin-coding-pro.md)
- [`avelin-coding-fast.md`](avelin-coding-fast.md)
- [Model Catalog](README.md)
- [API Reference](../api/index.md)

---

URL: https://avelin.ai/docs/models/avelin-agentic-fast

# avelin-agentic-fast

## Model Positioning

`avelin-agentic-fast` is AVELIN's cost-optimized agentic tier — delivering reliable tool-calling and agent execution at the lowest price point in the Agentic family. Built with AVELIN's Cross-Model MoE technology in our AI Lab, it fuses multiple model architectures into a single trained system tuned for high-volume agent operations.

**Context window:** 256K tokens input, 65K tokens output

**Pricing:** $0.285 input / $1.14 output per 1M tokens (cache read $0.0285, cache write $0.36) — see [Pricing](../pricing.md)

The right pick for agentic backends that need a cheap, fast, reliable workhorse behind the orchestration layer.

## Cross-Model MoE

Our Cross-Model MoE uses a prompt-vector classifier trained on 1 year of production data to predict the optimal specialist with 90% accuracy. It applies difficulty-based cascading (simple tasks → fast models, complex tasks → multi-expert fusion) and cost-aware routing (same quality, cheapest path). This is not a router — routers pick one model via if/else rules. We cascade across trained experts and synthesize output quality above any single expert.
Built with AVELIN's Cross-Model MoE technology in our AI Lab. Cross-Model MoE is a trained meta-model — not a router — that fuses knowledge from multiple model architectures into a single unified model. You call `avelin-agentic-fast` and get reliable agent execution at scale — the most cost-efficient model that meets your task requirements, without overpaying for unused reasoning capacity.

**Self-improving platform:** As new frontier models are released, AVELIN integrates them within 48 hours. Your `avelin-agentic-fast` endpoint gets more efficient over time — automatically, with zero code changes.

**Enterprise licensing:** AVELIN models are available for sovereign deployment on customer GPUs. [Request a POC](/poc#request) for enterprise licensing — run AVELIN on your own infrastructure with full data sovereignty.

**Why 256K context for high-volume agents:** When you're firing thousands of requests per workflow — classification, extraction, re-ranking, routing — you need speed and cost efficiency without sacrificing quality. 256K context is optimized for these high-throughput scenarios: enough for complex tool use and multi-step reasoning, but lean enough to keep per-token costs minimal. This is the workhorse that makes agent economics work at scale.

## Capabilities at a Glance

| Capability | Detail |
| --- | --- |
| Context window | **256K tokens** |
| Max output | **65K tokens** |
| Reasoning | off by default (enable per request) |
| Tools / function calling | yes (core strength) |
| Streaming | yes |

## Best-Fit Workloads

- High-volume agent swarms firing thousands of requests per workflow
- Bulk classification and retrieval re-ranking
- Prompt-routing and inner-loop tool calls
- Cost-sensitive agentic backends where every token counts

## When to Choose avelin-agentic-fast

- You're running high-volume agent orchestration
- Per-step cost matters more than peak reasoning depth
- The workload is standard agentic complexity
- You need reliable tool-calling at the lowest price point

For the most demanding agentic reasoning, use `avelin-agentic-ultra`.

## Example Business Requests

- "Classify these 10,000 support tickets by category."
- "Re-rank these 500 search results by relevance."
- "Extract entities from this batch of customer emails."
- "Route each of these queries to the appropriate downstream tool."

## Pricing

See [Pricing](../pricing.md) for complete pricing information. avelin-agentic-fast is the most cost-efficient agentic model in the AVELIN platform, optimized for high-volume workloads.

## Related

- [`avelin-agentic-pro.md`](avelin-agentic-pro.md)
- [`avelin-agentic-ultra.md`](avelin-agentic-ultra.md)
- [Model Catalog](README.md)
- [API Reference](../api/index.md)

---

URL: https://avelin.ai/docs/models/avelin-agentic-pro

# avelin-agentic-pro

> **Formerly `avelin-agentic`.** The legacy name keeps working as a server-side alias — no client change required.

## Model Positioning

`avelin-agentic-pro` is AVELIN's standard agentic tier — purpose-built for long-horizon tool-calling workflows and autonomous agents. Built with AVELIN's Cross-Model MoE technology in our AI Lab, it fuses multiple model architectures into a single trained system optimized for sustained multi-tool execution.

**Context window:** 256K tokens input, 65K tokens output

**Pricing:** $1.26 input / $3.96 output per 1M tokens (cache read $0.23, cache write $1.58) — see [Pricing](../pricing.md)

The standard agentic workhorse for tool-calling agents, regulated industries (healthcare, finance), and workloads where data sovereignty is required.

## Cross-Model MoE

Our Cross-Model MoE uses a prompt-vector classifier trained on 1 year of production data to predict the optimal specialist with 90% accuracy. It applies difficulty-based cascading (simple tasks → fast models, complex tasks → multi-expert fusion) and cost-aware routing (same quality, cheapest path). This is not a router — routers pick one model via if/else rules. We cascade across trained experts and synthesize output quality above any single expert.
Built with AVELIN's Cross-Model MoE technology in our AI Lab. Cross-Model MoE is a trained meta-model — not a router — that fuses knowledge from multiple model architectures into a single unified model. You call `avelin-agentic-pro` and get reliable autonomous execution based on tool-calling complexity, context requirements, and reasoning depth — without vendor lock-in.

**Self-improving platform:** As new frontier models are released, AVELIN integrates them within 48 hours. Your `avelin-agentic-pro` endpoint gets more capable over time — automatically, with zero code changes.

**Enterprise licensing:** AVELIN models are available for sovereign deployment on customer GPUs. [Request a POC](/poc#request) for enterprise licensing — run AVELIN on your own infrastructure with full data sovereignty.

**Why 256K context for agents:** Autonomous workflows require sustained reasoning across hundreds of tool calls, API interactions, and decision points. With 256K context, your agents can maintain coherent state, remember complex goals, and execute multi-step plans without fragmentation. This is the sweet spot for most production agent workflows — enough context for complexity, optimized for cost and speed.

## Capabilities at a Glance

| Capability | Detail |
| --- | --- |
| Context window | **256K tokens** |
| Max output | **65K tokens** |
| Reasoning | step-by-step (thinking) |
| Tools / function calling | yes (core strength) |
| Streaming | yes |

## Best-Fit Workloads

- Autonomous and semi-autonomous agents
- Multi-tool workflows (search, code, data, integrations)
- Task decomposition and sequential execution
- Orchestrated business automations via AVELIN-MCP

## When to Choose avelin-agentic-pro

- The task requires calling tools or APIs in a loop
- You need reliable planning and tool selection
- The workload is standard agentic complexity

For the most demanding agentic reasoning, use `avelin-agentic-ultra`.

## Example Business Requests

- "Plan and book a meeting, then email the attendees the agenda."
- "Research this account, then draft a tailored outreach sequence."
- "Pull the latest figures, build a table, and summarize the trend."

## Related

- [`avelin-agentic-ultra.md`](avelin-agentic-ultra.md)
- [Model Catalog](README.md)
- [API Reference](../api/index.md)

---

URL: https://avelin.ai/docs/models/avelin-agentic-ultra

# avelin-agentic-ultra

> **Formerly `avelin-agentic-high`.** The legacy name keeps working as a server-side alias — no client change required.

## Model Positioning

`avelin-agentic-ultra` is AVELIN's top-tier agentic reasoning model — the most capable model in the Agentic family for the most demanding tool-using and long-horizon tasks. Built with AVELIN's Cross-Model MoE technology in our AI Lab, it fuses multiple model architectures into a single trained system for complex autonomous workflows.

**Context window:** 256K tokens input, 65K tokens output

**Pricing:** $2.38 input / $11.88 output per 1M tokens (cache read $0.49, cache write $2.98) — see [Pricing](../pricing.md)

The strongest pick in the agentic family for security-sensitive tool-augmented workflows, deep research agents, and complex multi-step planning where quality matters more than per-step cost.

## Cross-Model MoE

Our Cross-Model MoE uses a prompt-vector classifier trained on 1 year of production data to predict the optimal specialist with 90% accuracy. It applies difficulty-based cascading (simple tasks → fast models, complex tasks → multi-expert fusion) and cost-aware routing (same quality, cheapest path). This is not a router — routers pick one model via if/else rules. We cascade across trained experts and synthesize output quality above any single expert.
Built with AVELIN's Cross-Model MoE technology in our AI Lab. Cross-Model MoE is a trained meta-model — not a router — that fuses knowledge from multiple model architectures into a single unified model. You call `avelin-agentic-ultra` and get maximized planning accuracy and tool-calling reliability — enterprise-grade autonomous intelligence without vendor lock-in.

**Self-improving platform:** As new frontier models are released, AVELIN integrates them within 48 hours. Your `avelin-agentic-ultra` endpoint gets more capable over time — automatically, with zero code changes.

**Enterprise licensing:** AVELIN models are available for sovereign deployment on customer GPUs. [Request a POC](/poc#request) for enterprise licensing — run AVELIN on your own infrastructure with full data sovereignty.

**Why 256K context for complex agents:** When your agent needs to orchestrate 100+ tool calls, maintain state across hours of autonomous work, or reason about complex systems with many moving parts, you need generous context. With 256K tokens, your agents can execute marathon workflows without losing coherence, forgetting goals, or fragmenting state. This is enterprise-grade autonomous intelligence.

## Capabilities at a Glance

| Capability | Detail |
| --- | --- |
| Context window | **256K tokens** |
| Max output | **65K tokens** |
| Reasoning | deep step-by-step (thinking) |
| Tools / function calling | yes (core strength) |
| Streaming | yes |

## Best-Fit Workloads

- Long-running agents that span many steps and tools
- Tasks over very large inputs (document sets, logs, data extracts)
- Complex orchestration where planning quality is critical
- High-stakes automations that must not lose context

## When to Choose avelin-agentic-ultra

- The agent must reason over a large amount of context at once
- Task complexity exceeds what `avelin-agentic-pro` handles comfortably
- Planning accuracy matters more than raw speed

## Example Business Requests

- "Audit this entire repository and propose a refactoring plan."
- "Coordinate a multi-step migration across these systems."
- "Analyze a quarter of meeting notes and produce an action roadmap."

## Related

- [`avelin-agentic-pro.md`](avelin-agentic-pro.md)
- [Model Catalog](README.md)
- [API Reference](../api/index.md)

# Earlier pages
> Pages that were on avelin.ai before the 3 Oct 2026 release. They keep their own text. Each is linked from the footer and from https://avelin.ai/resources, except the previous home page, which is linked from https://avelin.ai/resources and the sitemap only.
- [Business](https://avelin.ai/business): Three things: build models, serve them through an API, install them on your hardware.
- [Sovereign Corporate AI](https://avelin.ai/business/sovereign-corporate-ai): Frontier-class AI running inside your perimeter. On-prem or air-gapped. Your data never leaves.
- [AVELIN API](https://avelin.ai/business/avelin-api): One API for all our models. Automatic routing across 9 specialists.
- [Y-Ray Data](https://avelin.ai/business/y-ray-data): Sovereign AI data intelligence. Y-RAY connects, indexes, and reasons over your entire corporate knowledge base.
- [Use cases](https://avelin.ai/business/use-cases): Production-ready AI blueprints for regulated industries: Finance, Healthcare, Manufacturing, Government, and more.
- [AVELIN App](https://avelin.ai/private-gpt): AI workspace with chat, code, agents, tools, memory, and voice. In your browser or inside your perimeter.
- [Partners](https://avelin.ai/partners): AVELIN's strategic technology partners powering sovereign AI infrastructure worldwide.
- [Contact](https://avelin.ai/contact): Get in touch with AVELIN. Request a POC, ask about enterprise solutions, or connect with our team.
- [News](https://avelin.ai/news): Company news and announcements.
- [Earlier home page](https://avelin.ai/resources/earlier-home): An AI lab building sovereign, frontier-grade AI. Run it in our cloud like any API, or own it outright on your own hardware, on-prem or air-gapped.

# Use cases

> Blueprints by industry and eight priced workloads. Index page: https://avelin.ai/business/use-cases
## Finance
- PII-Safe Fraud Analysis (https://avelin.ai/business/use-cases/pii-safe-fraud-analysis): Intercept internal transaction logs, scrub Personally Identifiable Information (PII) locally, and route the anonymized payload to OpenAI or Claude for complex fraud detection without violating data privacy laws.
- M&A Due Diligence (https://avelin.ai/business/use-cases/ma-due-diligence): Parse and synthesize thousands of pages of confidential target-company financial records entirely locally, eliminating the risk of M&A data leakage.
- Algorithmic Fact-Checking (https://avelin.ai/business/use-cases/algorithmic-fact-checking): Cross-verify AI-generated earnings summaries against historical SEC filings using multi-agent debate to eliminate numerical hallucinations.
- Trading Strategy Backtesting (https://avelin.ai/business/use-cases/trading-strategy-backtesting): Automate the daily retraining of internal predictive models using end-of-day market states to keep proprietary trading algorithms continuously updated without manual engineering.
- Secure Core Banking Actions (https://avelin.ai/business/use-cases/secure-core-banking-actions): Connect Large Action Models (LAMs) to legacy mainframe databases using the Model Context Protocol (MCP) to execute secure, autonomous banking operations.
- Sovereign Wealth Management (https://avelin.ai/business/use-cases/sovereign-wealth-management): Deploy a fully offline, Kubernetes-native AI stack to analyze national asset data and build investment strategies without exposing financial intelligence to internet-connected infrastructure.
- Credit Risk Auditing (https://avelin.ai/business/use-cases/credit-risk-auditing): Generate transparent, mathematically verifiable audit logs of AI credit decisions using y-ray deep-trace to guarantee compliance with fair lending laws.
- Dynamic Claim Adjudication (https://avelin.ai/business/use-cases/dynamic-claim-adjudication): Deploy Large Action Models (LAMs) to autonomously extract unstructured data from claim PDFs, verify coverage rules, and execute API payouts for straight-through processing.
- High-Volatility Routing (https://avelin.ai/business/use-cases/high-volatility-routing): Guarantee zero downtime for critical financial AI applications by instantly routing compute away from degraded cloud LLM providers during major market events.
- Automated Compliance Red-Teaming (https://avelin.ai/business/use-cases/automated-compliance-red-teaming): Deploy an adversarial multi-agent workflow to autonomously draft and stress-test new financial products against SEC and FINRA regulations in real-time.
## Healthcare
- Drug Discovery IP Protection (https://avelin.ai/business/use-cases/drug-discovery-ip-protection): Run advanced molecular and chemical compound analysis entirely within your secure infrastructure, ensuring zero risk of IP leakage to third-party LLM providers.
- Clinical Trial Synthesis (https://avelin.ai/business/use-cases/clinical-trial-synthesis): Ingest, vectorize, and semantically search thousands of disjointed clinical trial PDFs to instantly map adverse effect correlations across multi-year studies.
- HIPAA-Compliant Patient Summaries (https://avelin.ai/business/use-cases/hipaa-patient-summaries): Automate clinical documentation by locally scrubbing Patient Health Information (PHI) before leveraging powerful external LLMs, ensuring strict HIPAA compliance and zero data leakage.
- Precision Medicine Retraining (https://avelin.ai/business/use-cases/precision-medicine-retraining): Automate the daily ingestion and fine-tuning of global genomic research to keep clinical diagnostic models operating on the absolute latest scientific consensus.
- Medical Billing LAM (https://avelin.ai/business/use-cases/medical-billing-lam): Automate medical coding and revenue cycle management by deploying Large Action Models (LAMs) to translate clinical notes into accurate ICD-10 codes and push them directly into internal billing systems.
- Hospital Resource Triage (https://avelin.ai/business/use-cases/hospital-resource-triage): Deploy lightweight, highly optimized inference models directly to hospital servers to guarantee AI-driven patient triage remains fully operational during complete internet or cloud outages.
- FDA Submission Verification (https://avelin.ai/business/use-cases/fda-submission-verification): Deploy a multi-agent swarm to autonomously cross-check and verify massive FDA submission documents against regulatory rubrics and internal clinical data to eliminate formatting errors and data contradictions.
- Compound Toxicity Prediction (https://avelin.ai/business/use-cases/compound-toxicity-prediction): Deploy multi-agent cognitive synthesis to cross-verify chemical interaction models, eliminating hallucination and generating highly reliable, AI-peer-reviewed toxicity predictions.
- Adverse Event Scraping (https://avelin.ai/business/use-cases/adverse-event-scraping): Deploy Large Action Models (LAMs) to autonomously scan global medical literature, extract adverse drug reactions, and format structured pharmacovigilance reports for immediate regulatory submission.
- Secure Medical RAG (https://avelin.ai/business/use-cases/secure-medical-rag): Deploy a localized Sovereign Corporate Brain that indexes fragmented hospital protocols and enforces strict Role-Based Access Control (RBAC), delivering instant, cited answers to clinical staff at the point of care.
## Oil, Gas & Heavy Manufacturing
- Offshore Rig Edge AI (https://avelin.ai/business/use-cases/offshore-rig-edge-ai): Deploy localized AI models directly onto physical offshore rigs to run real-time predictive maintenance in high-latency, completely disconnected environments.
- Seismic Data Analysis (https://avelin.ai/business/use-cases/seismic-data-analysis): Process highly classified geological survey data using powerful local AI inference models within a fully air-gapped environment to protect billion-dollar asset intelligence.
- Supply Chain LAM (https://avelin.ai/business/use-cases/supply-chain-lam): Deploy Large Action Models (LAMs) to continuously monitor enterprise inventory and autonomously execute secure procurement APIs, creating a self-healing supply chain that prevents production line stoppages.
- Engineering Blueprint Parsing (https://avelin.ai/business/use-cases/engineering-blueprint-parsing): Ingest, index, and query massive, unstructured Piping and Instrumentation Diagrams (P&IDs) using multimodal RAG to instantly retrieve exact engineering specifications.
- Refinery Zero-Downtime (https://avelin.ai/business/use-cases/refinery-zero-downtime): Ensure continuous AI-driven safety monitoring of high-risk refinery operations by deploying an automated failover architecture that instantly switches to localized backup models during cloud outages.
- Safety Incident Root Cause (https://avelin.ai/business/use-cases/safety-incident-root-cause): Deploy multi-agent cognitive synthesis to autonomously cross-examine contradictory sensor telemetry and human witness reports, isolating the factual root cause of industrial accidents.
- Daily Yield Optimization (https://avelin.ai/business/use-cases/daily-yield-optimization): Automate the daily retraining of refinery production models using the latest 24-hour yield data to continuously adapt to raw material fluctuations and maximize chemical output.
- Hazardous Material LAM (https://avelin.ai/business/use-cases/hazardous-material-lam): Deploy Large Action Models (LAMs) to continuously monitor pipeline telemetry and autonomously execute emergency shutdown protocols via MCP, preventing catastrophic industrial leaks.
- Vendor Contract Negotiation (https://avelin.ai/business/use-cases/vendor-contract-negotiation): Generate localized, AI-driven negotiation strategies by analyzing high-value supplier contracts and internal budgets securely without exposing supply chain intelligence to third-party cloud models.
- Emissions Compliance Logging (https://avelin.ai/business/use-cases/emissions-compliance-logging): Automate the calculation and reporting of facility emissions while generating mathematically verifiable, deep-trace audit logs to satisfy strict environmental regulations.
## Government, Defense & Public Sector
- Classified Intel Processing (https://avelin.ai/business/use-cases/classified-intel-processing): Process highly classified defense intelligence on a fully offline, Kubernetes-native AI platform, ensuring zero exposure to commercial clouds or internet-connected networks.
- Multi-Agency Data Fusion (https://avelin.ai/business/use-cases/multi-agency-data-fusion): Deploy an AI agent layer that securely queries heavily siloed law enforcement databases in real-time, enabling unified cross-agency intelligence without centralizing or moving the underlying classified data.
- Battlefield Edge AI (https://avelin.ai/business/use-cases/battlefield-edge-ai): Deploy hardware-optimized tactical AI models directly to forward-operating bases, delivering real-time strategic intelligence without relying on vulnerable or disconnected cloud networks.
- Citizen Request Anonymization (https://avelin.ai/business/use-cases/citizen-request-anonymization): Automate Freedom of Information Act (FOIA) redactions by locally scrubbing Personally Identifiable Information (PII) from public records before generating compliant, anonymized summaries.
- Threat Assessment Synthesis (https://avelin.ai/business/use-cases/threat-assessment-synthesis): Deploy multi-agent AI workflows to autonomously evaluate, debate, and verify the credibility of incoming cyber and physical threat alerts before escalating to human analysts.
- Procurement Fraud Deep-Search (https://avelin.ai/business/use-cases/procurement-fraud-deep-search): Ingest and analyze decades of unstructured municipal contracts using semantic deep-search to detect bidding anomalies, pricing manipulation, and hidden vendor relationships.
- Public Policy Simulation (https://avelin.ai/business/use-cases/public-policy-simulation): Continuously ingest regional economic data to retrain predictive AI models, enabling lawmakers to simulate the complex impact of proposed legislation on local employment, tax revenue, and GDP.
- Cyber Threat Hunting LAM (https://avelin.ai/business/use-cases/cyber-threat-hunting-lam): Deploy Large Action Models (LAMs) to continuously monitor network telemetry and autonomously execute millisecond-precision server isolation protocols via MCP to contain active cyber breaches.
- Legal Statute Drafting (https://avelin.ai/business/use-cases/legal-statute-drafting): Cross-check proposed legislation against centuries of existing statutes and constitutional law to automatically flag and resolve legal contradictions before a bill reaches the floor.
- Transparent Public Auditing (https://avelin.ai/business/use-cases/transparent-public-auditing): Generate mathematically verifiable audit logs of AI-driven government decisions to prove neutrality, eliminate bias, and satisfy strict public oversight requirements.
## Cross-Enterprise
- Enterprise LLM Gateway (https://avelin.ai/business/use-cases/enterprise-llm-gateway): Route all employee AI prompts through a sovereign interface to automatically scrub PII locally, enforce Role-Based Access Control, and completely eliminate shadow IT data leaks.
- Cost-Optimized Coding (https://avelin.ai/business/use-cases/cost-optimized-coding): Slash enterprise AI compute costs by 40-60% by using Intelligent Compute Routing to dynamically direct simple tasks to local open-source models and complex tasks to premium cloud APIs.
- Multi-Agent Contract Review (https://avelin.ai/business/use-cases/multi-agent-contract-review): Deploy an adversarial Red Team/Blue Team AI workflow to aggressively stress-test complex enterprise contracts, exposing hidden liabilities and legal loopholes that single LLMs miss.
- HR Onboarding LAM (https://avelin.ai/business/use-cases/hr-onboarding-lam): Deploy Large Action Models (LAMs) to autonomously read HR onboarding tickets, provision software licenses, and configure payroll systems via secure API execution, eliminating manual IT bottlenecks.
- Zero-Downtime Customer Support (https://avelin.ai/business/use-cases/zero-downtime-customer-support): Ensure continuous, 24/7 AI customer support by deploying an automated failover architecture that instantly reroutes traffic from degraded cloud APIs to localized backup models.
- Legacy Code Modernization (https://avelin.ai/business/use-cases/legacy-code-modernization): Deploy localized, air-gapped AI coding models to autonomously translate proprietary legacy codebases (e.g., COBOL) into modern stacks without exposing source code to public clouds.
- Daily Competitor Intel (https://avelin.ai/business/use-cases/daily-competitor-intel): Automate the daily ingestion of scraped competitor data to retrain internal sales enablement models, ensuring account executives always negotiate using real-time market intelligence.
- Secure Vendor Offboarding (https://avelin.ai/business/use-cases/secure-vendor-offboarding): Deploy Large Action Models (LAMs) to autonomously sweep enterprise directories and revoke external vendor credentials across all platforms instantly upon contract termination.
- Corporate IP Search (https://avelin.ai/business/use-cases/corporate-ip-search): Deploy a Sovereign Corporate Brain that indexes fragmented internal knowledge (wikis, Jira, Slack) to instantly provide employees with exact, cited answers instead of standard intranet keyword links.
- Hardware-Agnostic Experimentation (https://avelin.ai/business/use-cases/hardware-agnostic-experimentation): Decouple enterprise applications from underlying AI models using a universal abstraction layer, allowing engineering teams to instantly swap LLM providers without rewriting a single line of code.
## Priced workloads
- Ticket triage and bulk classification (https://avelin.ai/business/use-cases/ticket-triage): Label and draft replies for a high volume of short requests, and see what one ticket costs.
- Document extraction (https://avelin.ai/business/use-cases/document-extraction): Pull fields out of invoices, forms and agreements into structured data, and price each document.
- Agreement review against a checklist (https://avelin.ai/business/use-cases/agreement-review): Read long agreements against your own checklist and list the clauses that differ, with the clause text quoted.
- Code review and refactoring (https://avelin.ai/business/use-cases/code-review): Review pull requests and refactor across a whole repository in one context, with the repeated context cached.
- Research agent with web access (https://avelin.ai/business/use-cases/research-agent): Search, read and summarize sources with citations, and count the web requests and the tokens on one bill.
- Search and answers over your documents (https://avelin.ai/business/use-cases/document-search): Embed documents once, retrieve the passages for each question and answer from them, at list prices.
- Meeting transcripts and notes (https://avelin.ai/business/use-cases/meeting-notes): Transcribe recordings and turn them into notes. The audio is priced per minute and the notes per token.
- Workspace assistant for mail, calendar and tasks (https://avelin.ai/business/use-cases/workspace-assistant): An assistant that works in Gmail, Calendar, Contacts and Tasks, with the permissions each person grants.

### Ticket triage and bulk classification
URL: https://avelin.ai/business/use-cases/ticket-triage

# Ticket triage and bulk classification

Label and draft replies for a high volume of short requests, and see what one ticket costs.

### Why the bill is hard to read

Short, repetitive requests arrive in volume. Sending every one to the top model pays a flagship price for easy work. A bill per token does not show what one ticket costs.

### How AVELIN is used

Put each step on the tier that fits it. A fast tier labels the request, the pro tier drafts the reply, and only the cases the first pass could not settle reach the ultra tier or a person. In the API you can name the tier, or let the Decision Layer choose. A POC compares your split with what the Decision Layer would choose for the same tickets.

1. Label Send each request to avelin-agentic-fast with a fixed list of labels and ask for one label back.
2. Draft Send the labelled request and the matching help text to avelin-pro and ask for a draft reply.
3. Escalate Send requests the first pass could not label to avelin-ultra, or to a person.
4. Log Keep the label, the tier and the token counts for every request, so each ticket has its own cost.

### Tiers and list prices

USD per 1M tokens unless stated, list price as of 3 October 2026.

The models or endpoints this workload uses, with their list prices.
| Model or endpoint | Used for | List price |
| --- | --- | --- |
| avelin-agentic-fast | Label the request | $0.285 in, $1.14 out |
| avelin-pro | Draft the reply | $1.40 in, $4.40 out |
| avelin-ultra | Cases the first pass could not settle | $2.50 in, $12.50 out |

- Price list (https://avelin.ai/docs/pricing)

### What to count

Cost per task = the sum, over every call in the task, of (input tokens × input price + output tokens × output price) ÷ 1,000,000. Add cache write and cache read tokens at their own rates, and add any web requests at their per-request price.

- Calls per ticket, counted in a log of one real day.
- Input and output tokens per call.
- Which tier answered each call.
- How many tickets reach the ultra tier or a person.

Count this workload on your own runs, or request a POC and we price one workload at list price beside your current bill.

### FAQ: Questions about this use case

**What is "Ticket triage and bulk classification"?**
Label and draft replies for a high volume of short requests, and see what one ticket costs.

**Which tiers does it use?**
It uses avelin-agentic-fast, avelin-pro, avelin-ultra. Prices are in US dollars per 1M tokens, list price as of 3 October 2026.

**How do I test it on my own workload?**
A POC is a proof of concept on one of your workloads. It runs beside your current setup, with nothing in your production path, and ends in a cost-per-task report.
More: Request a POC (https://avelin.ai/poc#request)

Related: Document extraction (https://avelin.ai/business/use-cases/document-extraction); Agreement review against a checklist (https://avelin.ai/business/use-cases/agreement-review); Code review and refactoring (https://avelin.ai/business/use-cases/code-review); All use cases (https://avelin.ai/business/use-cases); Pricing (https://avelin.ai/pricing); Decision Layer (https://avelin.ai/decision-layer); Resources (https://avelin.ai/resources)

### Bring one workload. See its cost per task.

Start with a POC on one of your workloads.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

### Document extraction
URL: https://avelin.ai/business/use-cases/document-extraction

# Document extraction

Pull fields out of invoices, forms and agreements into structured data, and price each document.

### Why the bill is hard to read

A document is many pages and pages are many tokens. The bill depends on how long the documents are and on how often the same instructions are sent again.

### How AVELIN is used

Keep the field list and the instructions identical on every call. A repeated prompt prefix is billed at the cache read rate after the first call, so only the document itself pays the full input price. Start on the fast tier and send documents with missing fields to the pro tier.

1. Define Write the field list and the output format once, and use them as the start of every request.
2. Extract Send the document text, or page images, to avelin-fast. The fast tier accepts image input.
3. Check Validate the output against your own rules. Missing or doubtful fields go to avelin-pro.
4. Store Save the structured result together with the token counts for the document.

### Tiers and list prices

USD per 1M tokens unless stated, list price as of 3 October 2026.

The models or endpoints this workload uses, with their list prices.
| Model or endpoint | Used for | List price |
| --- | --- | --- |
| avelin-fast | First pass on every document | $0.30 in, $1.20 out. Cache read $0.03. |
| avelin-pro | Documents with missing or doubtful fields | $1.40 in, $4.40 out. Cache read $0.26. |

- Price list (https://avelin.ai/docs/pricing)

### What to count

Cost per task = the sum, over every call in the task, of (input tokens × input price + output tokens × output price) ÷ 1,000,000. Add cache write and cache read tokens at their own rates, and add any web requests at their per-request price.

- Pages per document and tokens per page, from a sample of your own documents.
- Tokens in the repeated prefix compared with tokens in the document.
- How many documents need the second pass.

Count this workload on your own runs, or request a POC and we price one workload at list price beside your current bill.

### FAQ: Questions about this use case

**What is "Document extraction"?**
Pull fields out of invoices, forms and agreements into structured data, and price each document.

**Which tiers does it use?**
It uses avelin-fast, avelin-pro. Prices are in US dollars per 1M tokens, list price as of 3 October 2026.

**How do I test it on my own workload?**
A POC is a proof of concept on one of your workloads. It runs beside your current setup, with nothing in your production path, and ends in a cost-per-task report.
More: Request a POC (https://avelin.ai/poc#request)

Related: Agreement review against a checklist (https://avelin.ai/business/use-cases/agreement-review); Code review and refactoring (https://avelin.ai/business/use-cases/code-review); Research agent with web access (https://avelin.ai/business/use-cases/research-agent); All use cases (https://avelin.ai/business/use-cases); Pricing (https://avelin.ai/pricing); Decision Layer (https://avelin.ai/decision-layer); Resources (https://avelin.ai/resources)

### Bring one workload. See its cost per task.

Start with a POC on one of your workloads.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

### Agreement review against a checklist
URL: https://avelin.ai/business/use-cases/agreement-review

# Agreement review against a checklist

Read long agreements against your own checklist and list the clauses that differ, with the clause text quoted.

### Why the bill is hard to read

Agreements are long and the checklist is the same every time. Sending the checklist again on every call is the largest avoidable cost.

### How AVELIN is used

Put the checklist first and the agreement after it, so the checklist is cached. The pro tier reads the agreement in its 256K-token context and lists the differences. Clauses marked hard go to avelin-ultra. A person decides every outcome. AVELIN gives no legal advice.

1. Prepare Write the checklist once and use it as a fixed prefix for every agreement.
2. Read avelin-pro compares each clause with the checklist and quotes the clause text it relied on.
3. Escalate Send the clauses marked hard to avelin-ultra for a second reading.
4. Decide A lawyer reviews the list and decides what to do with each item.

### Tiers and list prices

USD per 1M tokens unless stated, list price as of 3 October 2026.

The models or endpoints this workload uses, with their list prices.
| Model or endpoint | Used for | List price |
| --- | --- | --- |
| avelin-pro | Read the agreement against the checklist | $1.40 in, $4.40 out. Cache read $0.26. |
| avelin-ultra | Second reading of clauses marked hard | $2.50 in, $12.50 out. Cache read $0.52. |

- Price list (https://avelin.ai/docs/pricing)

### What to count

Cost per task = the sum, over every call in the task, of (input tokens × input price + output tokens × output price) ÷ 1,000,000. Add cache write and cache read tokens at their own rates, and add any web requests at their per-request price.

- Tokens in the checklist and in a typical agreement.
- Calls per agreement, including second readings.
- Cache read tokens compared with input tokens in the usage log.

Count this workload on your own runs, or request a POC and we price one workload at list price beside your current bill.

### FAQ: Questions about this use case

**What is "Agreement review against a checklist"?**
Read long agreements against your own checklist and list the clauses that differ, with the clause text quoted.

**Which tiers does it use?**
It uses avelin-pro, avelin-ultra. Prices are in US dollars per 1M tokens, list price as of 3 October 2026.

**How do I test it on my own workload?**
A POC is a proof of concept on one of your workloads. It runs beside your current setup, with nothing in your production path, and ends in a cost-per-task report.
More: Request a POC (https://avelin.ai/poc#request)

Related: Code review and refactoring (https://avelin.ai/business/use-cases/code-review); Research agent with web access (https://avelin.ai/business/use-cases/research-agent); Search and answers over your documents (https://avelin.ai/business/use-cases/document-search); All use cases (https://avelin.ai/business/use-cases); Pricing (https://avelin.ai/pricing); Decision Layer (https://avelin.ai/decision-layer); Resources (https://avelin.ai/resources)

### Bring one workload. See its cost per task.

Start with a POC on one of your workloads.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

### Code review and refactoring
URL: https://avelin.ai/business/use-cases/code-review

# Code review and refactoring

Review pull requests and refactor across a whole repository in one context, with the repeated context cached.

### Why the bill is hard to read

An agent that edits code makes many calls, and most of each call is the same repository context. A price per token hides how many calls one change needs.

### How AVELIN is used

The coding family has a 1M-token context, so a repository can fit in one request. Keep the repository context at the start of each call so it is cached. Use avelin-coding-fast for small fixes, avelin-coding-pro for review and multi-file changes, and avelin-coding-ultra for design and migration plans.

1. Load Put the repository map and the style guide at the start of the request.
2. Review avelin-coding-pro reads the diff and comments on it.
3. Change avelin-coding-fast makes small fixes. avelin-coding-pro handles multi-file edits.
4. Plan avelin-coding-ultra drafts the design for a large migration.

### Tiers and list prices

USD per 1M tokens unless stated, list price as of 3 October 2026.

The models or endpoints this workload uses, with their list prices.
| Model or endpoint | Used for | List price |
| --- | --- | --- |
| avelin-coding-fast | Small fixes | $0.30 in, $1.20 out. Cache read $0.03. |
| avelin-coding-pro | Review and multi-file changes | $1.40 in, $4.40 out. Cache read $0.26. |
| avelin-coding-ultra | Design and migration plans | $2.50 in, $12.50 out. Cache read $0.52. |

- Price list (https://avelin.ai/docs/pricing)

### What to count

Cost per task = the sum, over every call in the task, of (input tokens × input price + output tokens × output price) ÷ 1,000,000. Add cache write and cache read tokens at their own rates, and add any web requests at their per-request price.

- Calls per pull request or per fixed bug.
- Tokens in the repository context and in the diff.
- Cache read tokens compared with input tokens in the usage log.

Count this workload on your own runs, or request a POC and we price one workload at list price beside your current bill.

### FAQ: Questions about this use case

**What is "Code review and refactoring"?**
Review pull requests and refactor across a whole repository in one context, with the repeated context cached.

**Which tiers does it use?**
It uses avelin-coding-fast, avelin-coding-pro, avelin-coding-ultra. Prices are in US dollars per 1M tokens, list price as of 3 October 2026.

**How do I test it on my own workload?**
A POC is a proof of concept on one of your workloads. It runs beside your current setup, with nothing in your production path, and ends in a cost-per-task report.
More: Request a POC (https://avelin.ai/poc#request)

Related: Research agent with web access (https://avelin.ai/business/use-cases/research-agent); Search and answers over your documents (https://avelin.ai/business/use-cases/document-search); Meeting transcripts and notes (https://avelin.ai/business/use-cases/meeting-notes); All use cases (https://avelin.ai/business/use-cases); Pricing (https://avelin.ai/pricing); Decision Layer (https://avelin.ai/decision-layer); Resources (https://avelin.ai/resources)

### Bring one workload. See its cost per task.

Start with a POC on one of your workloads.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

### Research agent with web access
URL: https://avelin.ai/business/use-cases/research-agent

# Research agent with web access

Search, read and summarize sources with citations, and count the web requests and the tokens on one bill.

### Why the bill is hard to read

A research task mixes model calls and web requests. Counting only tokens misses the web half of the cost.

### How AVELIN is used

Use the search and scrape endpoints for sources and avelin-agentic-pro for planning and writing. Each web request has its own list price, so one task has two lines on the bill: model tokens and web requests.

1. Plan avelin-agentic-pro breaks the question into searches.
2. Search POST /v2/search for each query, at $0.005 per request.
3. Read POST /v2/scrape for each page, at $0.001 per request.
4. Write avelin-agentic-pro writes the answer and cites the URL of each source.

### Tiers and list prices

USD per 1M tokens unless stated, list price as of 3 October 2026.

The models or endpoints this workload uses, with their list prices.
| Model or endpoint | Used for | List price |
| --- | --- | --- |
| avelin-agentic-pro | Planning and writing | $1.26 in, $3.96 out |
| POST /v2/search | Web search | $0.005 per request |
| POST /v2/scrape | Read one page | $0.001 per request |

- Price list (https://avelin.ai/docs/pricing)

### What to count

Cost per task = the sum, over every call in the task, of (input tokens × input price + output tokens × output price) ÷ 1,000,000. Add cache write and cache read tokens at their own rates, and add any web requests at their per-request price.

- Searches and page reads per task.
- Tokens per planning call and per writing call.
- Total web requests multiplied by their list price, added to the token cost.

Count this workload on your own runs, or request a POC and we price one workload at list price beside your current bill.

### FAQ: Questions about this use case

**What is "Research agent with web access"?**
Search, read and summarize sources with citations, and count the web requests and the tokens on one bill.

**Which tiers does it use?**
It uses avelin-agentic-pro, POST /v2/search, POST /v2/scrape. Prices are in US dollars per 1M tokens, list price as of 3 October 2026.

**How do I test it on my own workload?**
A POC is a proof of concept on one of your workloads. It runs beside your current setup, with nothing in your production path, and ends in a cost-per-task report.
More: Request a POC (https://avelin.ai/poc#request)

Related: Search and answers over your documents (https://avelin.ai/business/use-cases/document-search); Meeting transcripts and notes (https://avelin.ai/business/use-cases/meeting-notes); Workspace assistant for mail, calendar and tasks (https://avelin.ai/business/use-cases/workspace-assistant); All use cases (https://avelin.ai/business/use-cases); Pricing (https://avelin.ai/pricing); Decision Layer (https://avelin.ai/decision-layer); Resources (https://avelin.ai/resources)

### Bring one workload. See its cost per task.

Start with a POC on one of your workloads.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

### Search and answers over your documents
URL: https://avelin.ai/business/use-cases/document-search

# Search and answers over your documents

Embed documents once, retrieve the passages for each question and answer from them, at list prices.

### Why the bill is hard to read

Embedding is cheap and done once. Answers are paid per question. The cost of an answer depends on how many passages you send with it.

### How AVELIN is used

Embed each document once with bge-m3. For each question, embed the question, fetch the closest passages and send them with the question to avelin-fast or avelin-pro. Y-Ray Data packages this with citations for teams that do not want to build it.

1. Embed Send each document to bge-m3 once and store the vectors.
2. Retrieve Embed the question and fetch the closest passages from your store.
3. Answer Send the question and the passages to avelin-fast, or to avelin-pro for harder questions.
4. Cite Return the passage and document each part of the answer came from.

### Tiers and list prices

USD per 1M tokens unless stated, list price as of 3 October 2026.

The models or endpoints this workload uses, with their list prices.
| Model or endpoint | Used for | List price |
| --- | --- | --- |
| bge-m3 | Embeddings, once per document | $0.02 per 1M tokens |
| avelin-fast | Answers to routine questions | $0.30 in, $1.20 out |
| avelin-pro | Answers to harder questions | $1.40 in, $4.40 out |

- Price list (https://avelin.ai/docs/pricing)

### What to count

Cost per task = the sum, over every call in the task, of (input tokens × input price + output tokens × output price) ÷ 1,000,000. Add cache write and cache read tokens at their own rates, and add any web requests at their per-request price.

- Tokens embedded per document, which you pay once.
- Passages and tokens sent with each question.
- Questions per day, which sets how much of the cost is chat.

Count this workload on your own runs, or request a POC and we price one workload at list price beside your current bill.

### FAQ: Questions about this use case

**What is "Search and answers over your documents"?**
Embed documents once, retrieve the passages for each question and answer from them, at list prices.

**Which tiers does it use?**
It uses bge-m3, avelin-fast, avelin-pro. Prices are in US dollars per 1M tokens, list price as of 3 October 2026.

**How do I test it on my own workload?**
A POC is a proof of concept on one of your workloads. It runs beside your current setup, with nothing in your production path, and ends in a cost-per-task report.
More: Request a POC (https://avelin.ai/poc#request)

Related: Meeting transcripts and notes (https://avelin.ai/business/use-cases/meeting-notes); Workspace assistant for mail, calendar and tasks (https://avelin.ai/business/use-cases/workspace-assistant); Ticket triage and bulk classification (https://avelin.ai/business/use-cases/ticket-triage); All use cases (https://avelin.ai/business/use-cases); Pricing (https://avelin.ai/pricing); Decision Layer (https://avelin.ai/decision-layer); Resources (https://avelin.ai/resources)

### Bring one workload. See its cost per task.

Start with a POC on one of your workloads.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

### Meeting transcripts and notes
URL: https://avelin.ai/business/use-cases/meeting-notes

# Meeting transcripts and notes

Transcribe recordings and turn them into notes. The audio is priced per minute and the notes per token.

### Why the bill is hard to read

Audio and text are billed in different units, so the cost of one meeting is two numbers added together.

### How AVELIN is used

whisper-large-v3 transcribes at $0.006 per minute. avelin-pro writes the notes from the transcript. Add the two. Transcription cost is the minutes of audio multiplied by the per-minute price.

1. Transcribe Send the recording to whisper-large-v3 and keep the transcript.
2. Summarize Send the transcript to avelin-pro and ask for the notes.
3. Extract Ask for a list of decisions and actions, each with the speaker named.
4. Store Save the notes with the minutes of audio and the token counts.

### Tiers and list prices

USD per 1M tokens unless stated, list price as of 3 October 2026.

The models or endpoints this workload uses, with their list prices.
| Model or endpoint | Used for | List price |
| --- | --- | --- |
| whisper-large-v3 | Speech to text | $0.006 per minute |
| avelin-pro | Notes from the transcript | $1.40 in, $4.40 out |

- Price list (https://avelin.ai/docs/pricing)

### What to count

Cost per task = the sum, over every call in the task, of (input tokens × input price + output tokens × output price) ÷ 1,000,000. Add cache write and cache read tokens at their own rates, and add any web requests at their per-request price.

- Minutes of audio per meeting.
- Transcript tokens sent for the notes.
- Notes tokens returned.

Count this workload on your own runs, or request a POC and we price one workload at list price beside your current bill.

### FAQ: Questions about this use case

**What is "Meeting transcripts and notes"?**
Transcribe recordings and turn them into notes. The audio is priced per minute and the notes per token.

**Which tiers does it use?**
It uses whisper-large-v3, avelin-pro. Prices are in US dollars per 1M tokens, list price as of 3 October 2026.

**How do I test it on my own workload?**
A POC is a proof of concept on one of your workloads. It runs beside your current setup, with nothing in your production path, and ends in a cost-per-task report.
More: Request a POC (https://avelin.ai/poc#request)

Related: Workspace assistant for mail, calendar and tasks (https://avelin.ai/business/use-cases/workspace-assistant); Ticket triage and bulk classification (https://avelin.ai/business/use-cases/ticket-triage); Document extraction (https://avelin.ai/business/use-cases/document-extraction); All use cases (https://avelin.ai/business/use-cases); Pricing (https://avelin.ai/pricing); Decision Layer (https://avelin.ai/decision-layer); Resources (https://avelin.ai/resources)

### Bring one workload. See its cost per task.

Start with a POC on one of your workloads.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

### Workspace assistant for mail, calendar and tasks
URL: https://avelin.ai/business/use-cases/workspace-assistant

# Workspace assistant for mail, calendar and tasks

An assistant that works in Gmail, Calendar, Contacts and Tasks, with the permissions each person grants.

### Why the bill is hard to read

An assistant that can act needs access, and access needs limits. Tools that can read everything by default are hard to approve.

### How AVELIN is used

AVELIN MCP connects the AVELIN App, or any MCP client, to Gmail, Calendar, Contacts and Tasks. Each connection is OAuth-secured and limited to the permissions the person grants. Give every API key its own budget and rate limit.

1. Connect The person connects an account and grants the permissions they choose.
2. Ask The person asks in plain language, for example to find a thread or move a meeting.
3. Act The model calls the tool that fits and returns what it did.
4. Review The person checks the result. Spend shows per key against its budget.

### Tiers and list prices

USD per 1M tokens unless stated, list price as of 3 October 2026.

The models or endpoints this workload uses, with their list prices.
| Model or endpoint | Used for | List price |
| --- | --- | --- |
| avelin-agentic-pro | Tool calls and multi-step requests | $1.26 in, $3.96 out |
| avelin-pro | Drafting and summaries | $1.40 in, $4.40 out |

- Price list (https://avelin.ai/docs/pricing)

### What to count

Cost per task = the sum, over every call in the task, of (input tokens × input price + output tokens × output price) ÷ 1,000,000. Add cache write and cache read tokens at their own rates, and add any web requests at their per-request price.

- Tool calls per request.
- Tokens per turn.
- Spend per key compared with its budget.

Count this workload on your own runs, or request a POC and we price one workload at list price beside your current bill.

### FAQ: Questions about this use case

**What is "Workspace assistant for mail, calendar and tasks"?**
An assistant that works in Gmail, Calendar, Contacts and Tasks, with the permissions each person grants.

**Which tiers does it use?**
It uses avelin-agentic-pro, avelin-pro. Prices are in US dollars per 1M tokens, list price as of 3 October 2026.

**How do I test it on my own workload?**
A POC is a proof of concept on one of your workloads. It runs beside your current setup, with nothing in your production path, and ends in a cost-per-task report.
More: Request a POC (https://avelin.ai/poc#request)

Related: Ticket triage and bulk classification (https://avelin.ai/business/use-cases/ticket-triage); Document extraction (https://avelin.ai/business/use-cases/document-extraction); Agreement review against a checklist (https://avelin.ai/business/use-cases/agreement-review); All use cases (https://avelin.ai/business/use-cases); Pricing (https://avelin.ai/pricing); Decision Layer (https://avelin.ai/decision-layer); Resources (https://avelin.ai/resources)

### Bring one workload. See its cost per task.

Start with a POC on one of your workloads.

- Request a POC (https://avelin.ai/poc#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

# News and research

Every post, newest first, under its title as published. Index page: https://avelin.ai/news

- [Web Summit 2026: AVELIN AI Among the Conference's Startups](https://avelin.ai/news/avelin-ai-takes-part-in-web-summit-2026-in-lisbon) (2026-09-28)
- [Sovereign AI at the Region's Top Fintech Event: AVELIN AI at Money20/20 Middle East 2026](https://avelin.ai/news/avelin-ai-at-money2020-middle-east-2026) (2026-09-14)
- [Venture Firm AltaIR Capital Accepts AVELIN AI into the AltaLab Program](https://avelin.ai/news/venture-firm-altair-capital-accepts-avelin-ai-into-the-altalab-program) (2026-09-11)
- [AVELIN AI Joins the Launchpad Program at Oraseya Capital](https://avelin.ai/news/avelin-ai-joins-the-launchpad-program-at-oraseya-capital) (2026-08-13)
- [Inc. Arabia Names AVELIN AI an AI Leader of Tomorrow at Its Dubai 2026 Awards](https://avelin.ai/news/avelin-ai-receives-the-award-in-the-ai-category-at-inc-arabias-leaders-of-tomorrow-awards-dubai-2026) (2026-07-30)
- [What is Sovereign AI?](https://avelin.ai/research/what-is-sovereign-ai) (2026-07-06)
- [The Agentic Multiplier: Why AI Costs are Rising as Prices Fall](https://avelin.ai/research/the-agentic-multiplier-why-ai-costs-are-rising-as-prices-fall) (2026-06-05)
- [The AI Commodity Era: Scaling Operations in a $1.3 Trillion Market](https://avelin.ai/research/the-ai-commodity-era-scaling-operations-in-a-13-trillion-market) (2026-06-03)
- [AVELIN AI Advances to DigitalOcean Partner Pod ISV Track to Scale Sovereign AI Globally](https://avelin.ai/news/avelin-ai-advances-to-digitalocean-partner-pod-isv-track-to-scale-sovereign-ai-globally) (2026-05-07)
- [AVELIN AI Receives Infrastructure Support from Lyceum Technology to Advance Sovereign AI](https://avelin.ai/news/avelin-ai-receives-infrastructure-support-from-lyceum-technology-to-advance-sovereign-ai) (2026-04-23)
- [AVELIN AI Secures AWS Infrastructure Backing via Incremental Workloads Program to Scale Sovereign AI](https://avelin.ai/news/avelin-ai-secures-aws-infrastructure-backing-via-incremental-workloads-program-to-scale-sovereign-ai) (2026-04-21)
- [AVELIN AI Joins Red Hat Partner Program to Scale Sovereign AI Deployments in Regulated Markets](https://avelin.ai/news/avelin-ai-joins-red-hat-partner-program-to-scale-sovereign-ai-deployments-in-regulated-markets) (2026-04-07)
- [AVELIN App-1.3.0 Release: A Massive Leap in Distributed AI Performance](https://avelin.ai/news/avelin-gpt-130-release-a-massive-leap-in-distributed-ai-performance) (2026-03-13)
- [AVELIN AI Joins DigitalOcean Hatch to Scale Sovereign AI for Regulated Enterprises](https://avelin.ai/news/avelin-ai-joins-digitalocean-hatch-to-scale-sovereign-ai-for-regulated-enterprises) (2026-03-11)
- [AVELIN AI Joins OVHcloud Startup Program](https://avelin.ai/news/avelin-ai-joins-ovhcloud-startup-program) (2026-03-05)
- [Engineering for Uncertainty: The Failsafe AI Architecture](https://avelin.ai/news/engineering-for-uncertainty-the-failsafe-ai-architecture) (2026-03-04)
- [AVELIN App Is Now Available Upon Request!](https://avelin.ai/news/avelin-gpt-is-now-available-upon-request) (2026-02-24)
- [The Architecture of Independence: An overview of AVELIN App](https://avelin.ai/news/the-architecture-of-independence-a-deep-dive-into-avelin-gpt) (2026-02-18)
- [AI Security Standards 2026: From On-Premise Infrastructure to SOC 2 Certification](https://avelin.ai/research/ai-security-standards-2026-from-on-premise-infrastructure-to-soc-2-certification) (2026-02-18)
- [The Trading Floor of Tomorrow: A.V.E.L.I.N. Reports from iFX EXPO Dubai 2026](https://avelin.ai/news/the-trading-floor-of-tomorrow-avelin-reports-from-ifx-expo-dubai-2026) (2026-02-13)
- [Observations from WHX Labs Dubai 2026: Why Global Healthcare Infrastructure is Still Waiting for the AI Age.](https://avelin.ai/news/observations-from-whx-labs-dubai-2026) (2026-02-12)
- [AVELIN Joins NVIDIA Inception to Power the Agent Economy](https://avelin.ai/news/avelin-joins-nvidia-inception-to-power-the-agent-economy) (2026-02-03)
- [Gateway to Growth: AVELIN Joins the Ignyte Ecosystem to Power the Next Generation of Sovereign Tech](https://avelin.ai/news/gateway-to-growth-avelin-joins-the-ignyte-ecosystem-to-power-the-next-generation-of-sovereign-tech) (2026-02-02)
- [The End of Hallucination: AVELIN Launches "Y-Ray Data" to Power the Sovereign Corporate Brain](https://avelin.ai/news/the-end-of-hallucination-avelin-launches-y-ray-data-to-power-the-sovereign-corporate-brain) (2026-01-30)
- [Sovereignty Meets Scale: AVELIN Deploys Strategic Infrastructure on Microsoft Azure](https://avelin.ai/news/sovereignty-meets-scale-avelin-deploys-strategic-infrastructure-on-microsoft-azure) (2026-01-30)
- [Sovereign AI as the New Global Currency: Economic Forecast 2026–2030](https://avelin.ai/research/sovereign-ai-as-the-new-global-currency-economic-forecast-20262030) (2026-01-28)
- [ABU DHABI | Machines Can Think Summit 2026](https://avelin.ai/news/yury-akinin-ceo-of-avelin-ai-at-the-machines-can-think-summit-in-abu-dhabi) (2026-01-27)
- [A.V.E.L.I.N. 2.0: The Next Evolution of Distributed AI](https://avelin.ai/news/avelin-2-launch) (2026-01-24)
- [AVELIN Joins Yellow Rocks Open Space Summit in Dubai](https://avelin.ai/news/yellow-rocks-summit) (2026-01-23)
- [AVELIN AI x ElevenLabs: We Are Giving Our Infrastructure a Voice](https://avelin.ai/news/elevenlabs-partnership) (2026-01-21)
- [A.V.E.L.I.N. Achieves SOC 2 Type II Certification](https://avelin.ai/news/soc2-certification) (2026-01-20)
- [Introducing the A.V.E.L.I.N. Avatar Marketplace](https://avelin.ai/news/avatar-marketplace) (2026-01-15)
- [Dubai Chamber of Digital Economy Supports AVELIN AI Through the Business in Dubai Program](https://avelin.ai/news/dubai-partnership) (2026-01-10)
- [The "Meta-Mind" Architecture: Solving the Enterprise Dilemma Through Hybrid Orchestration](https://avelin.ai/research/the-meta-mind-architecture-solving-the-enterprise-dilemma-through-hybrid-orchestration) (2026-01-08)
- [The Future of AI: Our 2026 Product Roadmap](https://avelin.ai/news/2026-roadmap) (2026-01-05)

# Legal

The legal pages are not copied here. Read them at these addresses.

- Data policy: https://avelin.ai/legal/data-policy
- API privacy policy: https://avelin.ai/legal/api-privacy
- API terms: https://avelin.ai/legal/api-terms
- Website privacy policy: https://avelin.ai/legal/privacy
- Website terms: https://avelin.ai/legal/terms
- App privacy: https://avelin.ai/legal/app-privacy
- App terms: https://avelin.ai/legal/app-terms
- MCP privacy: https://avelin.ai/legal/mcp-privacy
- MCP terms: https://avelin.ai/legal/mcp-terms