Mistral announced Responses API preview for agent builders

Mistral announced Responses API in public preview on October 28, 2026, for developers building agents with its hosted models. The API uses the OpenAI Responses format, but the application still has to retain conversation history and carry out any requested function calls.

What the preview accepts and returns

Requests can include text, images, or files. Image input requires a vision-capable model. Files can be supplied through an HTTPS URL, a base64 data URL, or an uploaded file ID. Mistral says the API works with all chat models it hosts, including Mistral Large 4.

A response can contain separate items for a model message, reasoning, and a function call. Structured output supports JSON mode and JSON Schema through text.format. Streaming includes text and function-argument events; applications should handle completed, failed, and incomplete responses, as well as error events. Vision and reasoning capabilities depend on the selected model.

For background on Mistral Large 4’s separate API preview, see NeoTeo’s earlier coverage of the model.

The application runs the function loop

Mistral Responses API is stateless. To continue a conversation, the application stores its history, including earlier output items, and resends them with the next request.

When the model returns a function call, it provides a function name and arguments. The application validates those arguments, runs the function in its own code, then sends the result back with the matching call_id. It also includes the relevant prior output items in the follow-up request. In other words, the model can request a tool call, but the application handles the work and feeds the result back into the conversation.

The documented response-creation endpoint is POST /v1/responses; its details are in Mistral’s API reference.

What format compatibility means for a migration

Mistral says Responses API uses the OpenAI Responses format, making that format relevant to teams already building with it. The shared format does not guarantee that every existing integration will work without code changes or that models behave identically. Developers still need to account for Mistral’s model-dependent capabilities and implement the application-side history and function handling.