# Evaluation

AVELIN models are evaluated continuously — and never self-graded.

## How we evaluate

- **Independent benchmarks.** Every model tier is measured against third-party, standardized evaluation suites covering reasoning, coding, and agentic tool use — the same public benchmarks used across the industry.
- **Production evaluation.** Beyond public benchmarks, we run our own continuous evaluation harness against live traffic patterns: quality, latency, cost per task, and reliability per model tier.
- **Continuous cycle.** New candidate models are evaluated before they serve a single production request, and deployed tiers are re-evaluated on a regular cycle. Routing weights update automatically when the data changes.

## Where the results live

Detailed, always-current evaluation results are published on our dedicated evaluation portal:

**[eval.avelin.ai](https://eval.avelin.ai)**

The portal shows per-tier results as they are produced, so numbers are never frozen into documentation that drifts out of date.

## What this means for you

- The model behind each AVELIN tier can improve at any time — your endpoint gets better with **zero code changes**
- A tier is only ever swapped to a backend that **meets or beats** the previous one on our evaluation suite
- Pricing and the API contract stay stable regardless of what serves the tier

## Related Documentation

- [Model Catalog](../models/README.md)
- [Pricing](../pricing.md)
- [How AVELIN Works](../systems/how-it-works.md)