Mistral's Responses API returns agent steps as separate items

Felix Laurent 29 Oct 2026 3 min read

ai agents

Mistral Responses API is now in public preview. Mistral AI announced it on Oct 28 as an API made for building agents on Mistral. We read the announcement and the new documentation to see what it changes for developers.

What Mistral announced

Mistral says each response returns the model's answer, its reasoning and its tool calls as separate items. The company says this lets an app handle every step more easily than with Chat Completions. That comparison is Mistral's own judgement.

The API works with all chat models hosted by Mistral, including Large 4. The quickstart uses mistral-large-4 with high reasoning effort. Mistral asks developers to send preview feedback to its support team. The docs show what those separate items look like.

What a response looks like

A response has an output list. It can hold message items with the answer text, reasoning items, and function_call items. Each function call carries a name, its arguments and a call_id.

Agent frameworks can read these typed items directly, alphasignal notes. They no longer need to pull tool calls out of one flattened assistant message. Input works the same way. A request can mix text, images for vision models and files, given as an HTTPS link, base64 data or an uploaded file ID.

The API also supports structured output with JSON mode or a JSON Schema, set through text.format. Reasoning is controlled with reasoning.effort, and plaintext reasoning depends on the model. What the API does not do is remember the conversation.

The app keeps the history

The documentation is direct about it: "This Responses API is stateless." The application keeps the conversation history and sends it in input with each request. That includes every output item from earlier responses, plus the new user message.

This is a difference from OpenAI's version of the format. OpenAI can store responses on its servers and chain them with previous_response_id. Mistral's preview docs mention neither option. Tool calls follow the same do-it-yourself pattern.

Tools run on your side

Function calling is client-executed, with support for namespaces. The model only asks for a call. The app must check the arguments, run the function and send back a function_call_output item with the matching call_id.

The Responses docs mention no built-in tools. Those stay in Mistral's separate Agents and Conversations API. That API keeps state across conversations and offers code execution, web search, image generation and a document library out of the box. Developers now choose between a stateful agent platform and a lighter, stateless loop they control.

Streaming and token counting

With streaming on, response.output_text.delta events carry the answer text as it arrives. Streams also include events for reasoning, function calls and the response lifecycle, and function arguments stream too.

The docs say to treat response.completed as the end signal. Apps should also handle response.failed, response.incomplete and error. A separate endpoint, POST /v1/responses/input_tokens, estimates input tokens before a call. Mistral says to send the same model, input, instructions and tools as the real request. The format itself comes from elsewhere.

Where the format comes from

OpenAI introduced the Responses API in March 2025 as its main interface for agents. In January 2026, Open Responses launched as an open, multi-provider specification based on that format. It is now a series of LF Projects. Its site lists OpenAI, Nvidia, AWS, Hugging Face, OpenRouter, Vercel, Ollama and vLLM among its backers. Mistral is not on that list.

Coverage framed the move as a way to switch from OpenAI more easily. Alphasignal's headline says developers can "ditch OpenAI instantly." Superpowerdaily calls it an OpenAI-format API for new apps and existing integrations. Neither outlet confirms that the OpenAI SDK works unchanged against Mistral's endpoint. At least one developer is waiting.

The maintainer of the open-source Kalamu project decided on Oct 28 to stay on Chat Completions for now. The issue cites the preview status and unconfirmed batch support for /v1/responses. It plans to revisit once the API is generally available.

Who should switch

So is the Responses API worth adopting now? For new agent code on Mistral models, it probably is. Typed items for answers, reasoning and tool calls remove a parsing step, and the format matches what many frameworks already speak. For teams with working Chat Completions code, the case is weaker. The API is in preview, it keeps no state and runs no tools for you, and batch support is unclear. The safest path is to try it on a new agent loop and wait for general availability before moving production traffic.