Skip to main content

Limits & supported fields

Every charge has to be calculable from the provider’s official usage. That’s why some endpoints, parameters and tools are deliberately not open to customer API keys. Requests that use them are rejected before they run, at no cost, with a message that names the field or endpoint responsible and says why.

Summary per endpoint

The model groups are explained in Chat Completions. Each endpoint’s page has the complete field table.

Endpoint Supported Rejected
/v1/chat/completions
OpenAI-format models
Text and images, function and custom tools, response_format, sampling parameters, streaming, and the other fields in this table. n other than 1, audio input and output, file parts, web_search_options and other server-side tools, prediction, service tiers other than the standard one, and any field not in the table.
/v1/chat/completions
Claude models
Text, images in user messages, function calling, temperature and top_p from 0 to 1, stop, response_format of type json_schema, thinking, streaming.seed, penalties, logit_bias, logprobs, reasoning_effort, response_format of type json_object, custom tools, cache_control, audio and file parts.
/v1/chat/completions
Text-and-tools models
Text, function calling, reasoning_effort, streaming.Images, sampling parameters, stop, response_format, thinking.
/v1/messages Text, images, documents, tools that you run, thinking, sampling parameters, stop_sequences, structured output, the 5-minute prompt cache, streaming. See Messages. Server-side tools (web search, web fetch, code execution, MCP), the 1-hour cache, other service tiers, anything that needs a beta header, content given by file ID, and models that don’t list the Anthropic format.
/v1/messages/count_tokensThe same body as /v1/messages. Free.Models that are not Claude models, server-side tools, images and documents given by URL.
/v1/responses Text and images, function and custom tools, reasoning, structured output, streaming. See Responses. store: true, previous_response_id, conversation, background, built-in tools, file and audio input, retrieving or cancelling a response, and models that list the Anthropic format.
/v1/embeddingsinput as a string, an array of strings, or an array of token IDs.Empty input.
/v1/images/*n from 1 to 10.n outside 1–10.
/v1/videos/generationsduration_seconds (required, max 3600), n from 1 to 10.A missing duration_seconds.

Not available

  • Server-side tools on every endpoint: web search, web fetch, code execution, file search, MCP servers, computer use and the like. Define a tool that your application runs instead.
  • Audio input and output, and files referenced by ID. Send file content as text, images or (on /v1/messages) base64 documents.
  • Stored responses: store: true, previous_response_id, conversation, background responses and GET, DELETE or cancel on /v1/responses/{id}.
  • Other processing tiers: priority, flex and fast modes, a chosen inference region, and the 1-hour prompt cache.
  • Any endpoint not listed in the Introduction, such as audio, files, batch or fine-tuning.

Time limits

  • 90 seconds to the first response. A reply without streaming has to be complete within 90 seconds; a stream only has to start within that time. Use "stream": true for a long generation. Image and video requests have to finish within the same limit.
  • A request that passes the limit is answered with 408 if it had not reached the model, or with 422 timeout_error if it had. See Errors & status codes.
  • 300 seconds of silence. Once a stream has started, the model may stay silent for up to 300 seconds at a time. After that the stream is cut, and you are charged for the usage reported up to then.
  • 30 minutes per stream in total.
  • While a stream is quiet, a comment line : keep-alive is sent about every 15 seconds so that no proxy closes the connection. The official SDKs ignore comment lines; a parser you write yourself must skip every line that starts with :.

Other limits

  • max_tokens (or max_output_tokens) must be greater than 0 and no larger than the model’s context limit.
  • A request body may be up to 16 MB; a larger one is rejected with 413.
  • Each API key may send 600 requests per minute and have 16 requests in progress at once; an account may have 32 in progress across all its keys. Beyond that, requests get 429 with a Retry-After header. Token counting counts towards the same limits.
  • A request may define up to 128 tools on /v1/responses and for text-and-tools models.
  • Every request has a maximum hold. See Billing & balance.
  • An account can have up to 20 active API keys.

Start with any balance,
however small.

Create a free accountSign in AI API · pay as you go