Skip to main content

​
Model catalog

GET /v1/models is the live catalog. Every entry carries the model’s name, context window, accepted inputs and your price. GET /v1/models/{model} returns a single entry.
The catalog changes as models are added or retired. Read it from the API instead of hard-coding a list.
A model name has the form vendor/model, for example openai/gpt-4o-mini. Case and separators do not matter: anthropic/claude-haiku-4.5 and anthropic/claude-haiku-4-5 are the same model.
​
string
required
The value to put in model.
​
string
Display name.
​
string
required
The model’s vendor.
​
string
text — served by /v1/chat/completions and /v1/responses; image, video, music, audio — by the matching /v1/{kind}/jobs.
​
integer
Context window in tokens (prompt plus answer).
​
integer
Largest answer the model can produce, when known.
​
string[]
text, image, audio, video, file.
​
object
required
Your price in USD. Text models: input_per_mtok, output_per_mtok and cached_input_per_mtok (per million tokens). Media models: unit (image, video, second, request, 1k_characters) and per_unit for the cheapest configuration.
Fields a model has no data for are omitted rather than sent as zero.

​
Cost

Requests are charged in US dollars, counted in micro-dollars: 1 µ==0.000001. Text:
  • A price per million tokens is also the price per token in µ$.
  • Token counts come from the model’s own usage report, returned to you unchanged.
  • Any request that consumed tokens costs at least 1 µ$; the total is rounded up.
  • Hidden reasoning tokens are billed as output tokens.
Media: a job is charged once, when it completes — see cost. estimated_cost is the forecast at start. Duration, resolution, sound and tier raise the price. A failed job costs nothing.

​
Balance holds

Before a request runs, its largest possible cost is held against your balance, so parallel requests cannot together spend more than you have.
  • Text — the prompt plus max_tokens of answer (4096 if unset). When the request ends, the hold is replaced by the actual cost.
  • Media — the job’s cost. The first time a configuration is used (a new model, a first 4K clip), the hold carries a margin. A failed job releases its hold at once.
If the hold does not fit in your free balance, the request is refused with 402 insufficient_balance before any work begins. meta gives required_micro_usd, balance_micro_usd and held_micro_usd.
When your balance is small, always set max_tokens.

​
Reasoning models

o-series, DeepSeek R1 and thinking variants, Gemini Pro and Claude thinking spend part of max_tokens on hidden reasoning. Give them max_tokens of 2000 or more, or the answer may be cut off with finish_reason: "length".