Decision Layer
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.
The top model is the default. Most tasks do not need it.
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.
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.
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.
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.
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.
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.
What it includes.
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.
Quality analysis
Answers are scored, so you can see whether a cheaper model held up on your own tasks.
Dangerous-request detection
The Decision Layer spots requests that look dangerous before they reach a model.
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.
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.
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.
Our Decision Layer.
It reads each task and picks a model from the providers you connect, next to the choice you make today.
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.
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%.
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.
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.
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.
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.
Prove it on your own tasks.
Start with a POC on your own tasks and see your cost per task before anything changes.