Get started

Compatibility

What each of the three formats accepts, at a glance, and the few requests that are refused.

On this page

One key, three formats

One key, the same models, three reply formats: Chat completions, Responses and Messages. Use the one your code or SDK already uses. The differences are in the request, not in what a model can do. A model built for decisions is the exception: it answers its own endpoint only.

This page shows what each format accepts, so you can tell quickly whether a request works as it is. The fields, the reply and the refusals in full are on that format’s own page.

What each format accepts

RequestChat completionsResponsesMessages
A streamed replyYesYesYes
Function tools, with toolsYesYesYes
Choosing a tool with tool_choiceYesYesYes
Text from an imageimage_url partsinput_image with a URLimage blocks, URL or base64
Reasoningreasoning_effortreasoningthinking
Structured outputresponse_formattext.formatNo
Temperature, top_p and stopYesYesYes
More than one choice in a replyNo——
State kept between requestsNoNoNo
Decision questions, at /v1/systemone———

A refusal is never quiet: the request answers 400, naming the field in param. A client that sends something this API doesn’t do hears about it on the first call, not in the reply.

Where the habits differ

Before you point an existing client at this API, there are three habits to know.

  • Nothing is stored. There is no previous_response_id, no conversation and no stored reply: send the earlier turns in input or messages on every call. store is accepted and ignored.
  • A model may be called in all three formats, but a field that only one format has stays there. Structured output lives on Chat completions and Responses; Messages refuses it, and its own approach is a tool the model has to call.
  • Model IDs are this API’s, not another API’s. List models gives the IDs this key can call, which is also the list an SDK asking for models will see.

What isn’t there

WhatWhat happens
Files, batches, assistants, threads, embeddings, audio and image generationNot served. Only the three reply formats, their model lists and count tokens exist.
Long-running or background requestsRefused with a 400. A streamed reply is the way to wait for a long one.
Multiple choices in one requestRefused with a 400 on Chat completions. Send separate requests.

Every format’s page lists what it accepts and what each refusal looks like: Chat completions, Responses, Messages.

Requests are standard HTTPS with your key in a header, so an SDK needs only the base URL and the key: see OpenAI and Anthropic SDKs. Error codes are on Errors.