methodology

How we choose models, prices and tests

A ranking or price is only useful if you know how it was produced. This page explains the rules behind the catalog, the rate sheet and the tests we plan to publish.

Last updated: 2026-10-23

Catalog inclusion

A model is listed when it is available from the upstream catalog, has a usable text output, is not retired or excluded by slug, and has a resolvable input and output price. Uncensored models are identified by model name, maker and catalog metadata. The label is a routing aid, not an independent safety or quality certification.

Pricing

Supplier list price is the head of the routing waterfall: the first default provider that is host-priced and routable. Retail rates are computed from that cost, with an uncensored markup and a frontier pass-through. Admin overrides can change a rate. The model page shows the effective retail rate, not the supplier cost. Per-request charges use the provider's token counts and are recorded as metadata.

What a price does not tell you

Token counts differ by tokenizer. A model with a lower per-token rate can be more expensive for the same text. Latency, retries and output length also affect the real cost of a workload. Use the calculator for a rough estimate, then measure your own requests.

Benchmark plan

We do not publish vendor-reported refusal numbers as if they were our own. The planned benchmark uses a fixed prompt set across a sample of enabled models, with repeated runs, and reports refusal rate, completion rate, latency and a small coherence check. Results will be published with the prompt set, model IDs, date and run conditions so they can be reproduced. Until a run is complete, no result is shown.

Status and corrections

The status page reports observed success rate and average latency from real customer requests over 24 hours. It is not synthetic uptime monitoring. If a model page, price or comparison is wrong, report it with the URL and the observed value. We correct material errors and note the change date on the page.

Independence

Comparisons with OpenRouter, Venice AI and other services are directional and based on publicly available information at the time of writing. We are not affiliated with those services. Verify their current terms before relying on a comparison for a purchase decision.

Questions

How are uncensored models identified?

By model name, maker and catalog metadata. The label is a routing aid, not an independent quality or safety certification.

Do you publish benchmark results?

Only after a reproducible run is complete. Until then, the methodology page describes the plan rather than showing vendor numbers.

How can I report an error?

Use the contact channel on the status page and include the URL and the observed value.

Create account - no emailSee pricingBrowse models