Mistral opens a Responses API in public preview, using OpenAI's request format

Mistral has introduced a Responses API in public preview. It follows the OpenAI Responses format, so code written for that format can be pointed at https://api.mistral.ai/v1/responses with fewer changes. Mistral says each response returns “the model’s answer, its reasoning and its tool calls as separate items”, and that it works with all chat models Mistral hosts, including Mistral Large 4.

What works in the preview

The Responses API guide lists what you can use today:

• text input and output, as strings or message arrays;
• image input (URLs or base64) with a vision-capable model, and file input (HTTPS URLs, base64 or uploaded file IDs);
• structured output with JSON mode or JSON Schema via text.format;
• streaming events for text and function arguments;
• function calling with functions your own code executes, including namespaces;
• reasoning.effort and plaintext reasoning, where the model supports it;
• input token counting at POST /v1/responses/input_tokens.

What it does not do yet

The API is stateless. Your application keeps the conversation history and sends it back in input on every request, including the output items from earlier responses. The API reference says conversation must be omitted or null and background must be false, so there is no server-side thread and no background mode. max_tool_calls is accepted for compatibility but ignored.

Tools are the other gap. The preview lists only client-executed functions, with no hosted tools such as web search or a code runner: the model asks, your code runs the function and sends back the result.

Who should try it

If you already use a Responses-style client and want to try Mistral models, this lowers the switching cost. If you depend on server-side conversation state or built-in tools, stay where you are for now and watch the preview. Mistral’s own examples use mistral-large-4 with high reasoning effort; if you switch models, the guide tells you to pick a reasoning effort that model supports.