Skip to main content
GET
No authentication required.

​
The three axes

Every model is described on three independent axes. They answer different questions, and keeping them apart is what makes the catalog filterable. The distinction that matters: image is a type; generate-images is a use case. One describes the model, the other describes your job.

​
type — 10 values

language · image · video · embedding · reranking · transcription · audio-understanding · realtime · speech · decisions type is the product line you buy the model under — which endpoint serves it and which unit it bills in — not simply what it emits. That is why whisper-v3-turbo, which returns text, is transcription: you reach it at POST /v1/audio/transcriptions and it bills per minute of audio, not per token. realtime and audio-understanding list no models today: realtime sessions are paused, and audio understanding was retired on 2026-10-08, so ?type=realtime and ?type=audio-understanding return an empty list. The audio types split by what you are asking for: transcription writes the words down, audio-understanding answers questions about the sound itself, speech generates audio, and realtime is the bidirectional socket. decisions is the same rule once more: a decisions model takes text and returns a choice with a confidence, never prose, so it is not a language model — you reach it at POST /v1/decisions and it bills per input token.

​
tags — 12 values

tool-use · reasoning · vision · implicit-caching · explicit-caching · file-input · web-search · image-generation · video-generation · websocket-realtime · websocket-transcription · fast A tag is only set when it is objectively true of the model’s own spec. Three of them — fast, implicit-caching, explicit-caching — describe measured behaviour of a serving route rather than the model, so they are currently unset on every model rather than asserted without evidence.

​
use_cases — 14 values

write-code · build-agents · chat · write-content · translate · research-web · analyse-documents · solve-math · generate-images · edit-images · generate-video · transcribe-audio · generate-speech · search-documents

​
Filtering

All filters are query params. Comma-separated values combine as AND, except type — a model has exactly one type, so listing several means OR. Combine tools and structured_outputs with type=language. On their own they also return image, embedding, speech and transcription models whose row says supports_tools: false. The same goes for max_input_price: models billed per image, minute or character report an input price of 0, so they pass any ceiling. creator is who made the model — the value each row carries as owned_by. ?owned_by= is accepted as the same filter, so you can filter by exactly what you just read in the response. Filters compose:
The filters block in the response echoes back exactly what was applied, so you can confirm a param was understood rather than ignored.
recommended_for is deprecated. It is now an alias of use_cases and still works, including the old vocabulary (coding, agent, general, image, …), which is translated to its use_cases equivalent. Prefer use_cases for new integrations. Nothing has been removed — recommended_for remains in every response.

​
free_tier — what a new key can call

Every row carries free_tier, a boolean. true means the model is in the free tier: the $0.50 signup credit, added once you verify your email, spends on it, so an account that has never added credits can call it. false means it needs a top-up first, and a request from a never-funded key is refused with 402 model_requires_paid before anything is charged.
cost_tier and tier are editorial labels — “cheap”, “balanced”, “premium” — and nothing enforces them. free_tier is the rule the API applies. It is derived from each model’s published price rather than kept as a list of names, so it follows the catalog: a model whose price moves across the line changes side the same day. The line itself, and what it is set to, is on the pricing page. GET /v1/models/recommend carries the field on every model it names, so a recommendation can be filtered the same way.

​
Response

The row above is abridged: live rows also carry created, weights_source, tier, capability, release_date and output_limit_source.
A model can be served by more than one piece of infrastructure. context_window, max_output_tokens and the supports_* flags describe what is achievable for that model — the largest window reachable, and every capability at least one route can serve. Requests are matched to a route that can handle them, so a call that needs a capability never lands somewhere without it.