Mistral Opens Responses API in Public Preview for Agents
Mistral's Responses API is in public preview for agent builders, returning answers, reasoning and tool calls as separate items, in OpenAI Responses format.
Anyone who has built an agent knows the annoyance: a single model call returns an answer, some reasoning and a tool call to run, all packed into one message that's painful to parse. Mistral said on Wednesday that its Responses API, designed for building agents, is now in public preview, and its approach is to pull those pieces apart.
One response, several separate items
According to Mistral, each response returns the model's answer, its reasoning and its tool calls as separate items, so an application can handle each step on its own rather than carving it out of one big output. Compared with the Chat Completions interface, that suits agents that run many steps and call tools repeatedly.
The docs show it uses the OpenAI Responses format, so applications already built on that format should find migration easier. The endpoint is POST /v1/responses, taking a model, an input and optional reasoning settings and tools. Mistral says it works with all chat models it hosts, including Large 4.
What the preview supports
The preview covers text input and output, image and file input, JSON mode and JSON Schema structured output, streaming, client-executed function calls, model-dependent reasoning effort, and an endpoint for counting input tokens.
One thing to note is that it is stateless. Your application keeps the conversation history and must send earlier output items back with each request. Mistral also has separate Agents and Conversations APIs that store the conversation for you, so the two approaches differ in where the state lives.
Why it matters
An API format may look like a detail, but it shapes how costly it is to move around the ecosystem. The Responses format originated with OpenAI. By staying compatible, Mistral lets developers change models with fewer code changes, and makes its models easier to plug into tools and apps already written for that format.
It is still a public preview, and Mistral's docs invite feedback through its support team. Developers curious about it can take an existing Responses integration, swap in a Mistral model, and see whether handling reasoning items and tool calls feels smoother before deciding whether to migrate.