Where these numbers come from.

The rules

1. Every price is read off the provider's own pricing page. Not a press release, not another aggregator, not a model's own answer about itself.
2. Every price carries the date it was read and a link to the source.Both appear on the model's page so you can check our work in one click.
3. An unverified price is not published. Models without a confirmed source are held out of the board entirely rather than shown with a caveat. A caveat nobody reads is not a safeguard.
4. Prices are re-checked every 45 days. Anything older is flagged here and in the build.
5. No affiliate links and no paid placement anywhere on the price board or in the tool directory. If that ever changes it will be disclosed on this page first.

What the numbers mean

All prices are US dollars per million tokens, at standard list rates on the provider's first-party API. Where a provider runs an introductory or promotional rate, we show the rate you would actually pay today and the list price it reverts to, with the date.

Cache-read rates are the price of re-reading a cached prompt prefix. Where a provider expresses this as a multiple of the base input rate rather than a separate number, we compute it and say so in the source notes.

Third-party platform pricing — Amazon Bedrock, Google Vertex AI, Microsoft Foundry — is set by those platforms and is not reflected here unless stated. Check the platform's own pricing page.

What we do not cover

Batch discounts, committed-use and enterprise agreements, fine-tuning, embeddings, image and audio pricing, and per-request minimums are all out of scope. So are benchmark scores: plenty of places rank models on quality, and doing it badly is worse than not doing it.