Every price and spec on this site carries a source and a date. This page explains where the data comes from, how often it is checked, and what happens when we cannot verify something.
Where the data comes from
- Public machine-readable pricing feeds — including the LiteLLM pricing dataset, refreshed on a schedule. These give us API rates per token, per image and per second.
- Provider pricing and docs pages — for everything a feed does not cover: subscriptions, credit packs, free-tier limits, licences.
- Manual entries and community submissions — accepted only with a URL we can re-check.
What we never do
- We do not estimate a price we cannot source. Unknown shows as Not disclosed.
- We do not mix editorial opinion with benchmark data — our own ratings are labelled as such, third-party scores link to their source.
- We do not silently overwrite good data with a bad parse. Large or surprising changes go to human review first.
Cost-per-task estimates
Headline “cost per task” numbers are calculated from published rates using fixed, editable assumptions: a 1,000-word article is 1,500 input + 500 output tokens; an image is one 1024 px generation — counted as 1,056 output image tokens where the provider bills that way (OpenAI’s published figure for a medium-quality image); a video clip is 10 seconds; a minute of voiceover is 60 seconds of speech, or 750 characters where the provider bills per character. Open-weight models you host yourself show as Free, self-hosted — the marginal cost is zero, the hardware is yours. Units we do not model (seats, credit packs) show Not disclosed rather than an invented number. These are estimates — the provider’s own calculator is the final word.
Corrections
Spotted something wrong? Every model page has a Suggest an edit link, and providers can submit or claim a listing. Corrections with a source are prioritised.
Browse the results: all AI models · recent pricing changes.