What qualifies
A provider must run the model remotely and expose a callable API. We include text generation, embeddings, reranking, speech, image, safety, and user-deployed model endpoints when hosted compute is genuinely covered by the offer.
What does not
- A free web chat with no developer API
- Free weights that you must host yourself
- A free gateway that still bills the upstream model
- Sandbox responses that are mocked rather than inferred
- Stale launch posts contradicted by current billing docs
How we label free
A recurring or indefinite allowance measured in requests, tokens, or compute.
Named model IDs priced at zero inside a changing provider or router catalog.
A dollar-like balance that refreshes on a published schedule.
Free evaluation or research access that may forbid production or commercial use.
Volunteer or donation-supported inference with variable availability.
A signup credit or fixed trial window. Useful, but never presented as permanently free.
Evidence order
Live official catalogs come first, followed by current pricing, billing and limits documentation, product pages, changelogs, official repositories, company announcements, and then reputable corroboration. Search snippets and third-party trackers are leads—not proof.
Know the limit
Most model endpoints require an account, so a public catalog cannot prove an authenticated generation or a zero-dollar billing result. Those gaps are stated plainly, and conflicting sources stay on the watchlist.