Mistral opens Responses API in public preview, separating reasoning and tool calls for agents

01In brief

The new endpoint is for building agents on Mistral and is stateless: your application must store the conversation history and resend it with every request (API docs).

02What's happening

Mistral has added the POST /v1/responses endpoint, which is currently in public preview (@MistralAI). According to the official API docs, each returned item is labeled by type: message is the answer, reasoning is the reasoning, and function_call is the model's request to run a tool. The endpoint also supports function tools, parallel tool calls, streaming, and JSON Schema output. Whether reasoning effort can be adjusted depends on the model.

The preview has several limitations: you cannot continue from the previous response, background execution is not supported, and inputs that exceed the model's context length are not automatically truncated; the max_tool_calls parameter is ignored. The docs do not list supported models; the example uses mistral-large-4. This is based on the official announcement and docs (🗣 self-reported); we have not tested actual API calls ourselves.

03What this means for you

Developers building agents with Mistral can distinguish reasoning from tool calls by item type, without parsing the text themselves. Tools still need to be run in your application, which then returns the results. This is still a preview, so confirm which capabilities your model supports before using it in production.

More briefs →