# 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)

URL: https://avelin.ai/method
Updated: 3 Oct 2026