Mistral has opened a public preview of its Responses application programming interface for agent applications. Model answers, reasoning and tool calls arrive as separate items rather than one combined message. Developers can submit text, images and files, while their own applications execute requested functions and supply the conversation history needed for the next turn.
AI developer Mistral opened its Responses application programming interface, or API, in public preview on October 28. The interface on La Plateforme, its model-hosting service, returns reasoning, requested tool calls and message answers as separate output items. Applications can handle each type individually. It uses the Responses format familiar from OpenAI integrations and offers a new interface to models already hosted by Mistral.[1], [2]
Mistral’s implementation does not retain the conversation between requests. The application keeps previous items and supplies the history needed for the next turn. Function execution also remains with the client: when a model asks to call a tool, the application runs the function and returns its result to the model. These are successive steps in the same interaction. The new output format separates their contents while leaving responsibility for history and execution with the developer’s software.[1]
The preview accepts text, images and files. It also supports streaming and structured output in JSON, a structured data format applications can process by named fields. A separate input-token endpoint counts the units of input before a request. Developers can use Python and TypeScript tools to work with the interface. Controls for reasoning effort depend on the selected model’s capabilities. The availability stage remains a public preview, with the chosen model determining which inputs and controls an application can use.[1]
