Models and list rates

How to read these prices

A list rate is what a lab or a router charges for API access, per million tokens. It is not what a Claude Pro, ChatGPT Plus or Cursor subscription costs you, and it is not a bill from tokenstat. It is the yardstick that lets you compare one model against another, and see what a session would have cost if you had paid for every token.

Output costs more, but cache reads decide the total

Output tokens usually cost several times more than input, so the output column looks like the one that matters. In a coding agent it rarely is. Every turn re-sends the conversation so far, and the vendors serve most of it from a cache at a fraction of the input rate. On a typical machine cache reads are well over 90% of all tokens counted, so a model's cache-read rate often moves your real cost more than its headline price. Why cache reads dominate.

The same model, many prices

A popular model is resold by gateways and hosts, each at its own rate, and the cheapest listing is not always the lab's. The Offers column counts those listings. Each model page shows them side by side, with the lab's own rate first when we have checked it against the lab's pricing page.

Where the numbers come from

The catalog merges LiteLLM, OpenRouter and models.dev with a short list of rates copied by hand from the labs, rebuilt every hour. A feed that suddenly moves a price by more than half is held back until it holds up, because a broken feed is more common than a real overnight price cut. How the catalog is built.