Mistral AI announced on Wednesday, 28 October 2026, on its social media, the public preview deployment of its Responses API. This interface, designed to facilitate the creation of agents based on artificial intelligence, adopts the format introduced by OpenAI with its Responses API, launched in 2025.
An API that details each step of the model’s work
“For each call, the API returns the text generated by the model, its reasoning and its tool calls as separate elements,” explains Mistral AI. According to the company, this is what distinguishes the Responses API from Chat Completions, launched several months ago, which processes a list of messages and produces a single output that can blend text, citations and tool calls.
Mistral AI justifies this choice by the evolution of use cases: “Many new tools for agents rely solely on this format.” The API works, according to the company, “with all its text models, including Mistral Large 4.” The company also takes the opportunity to add to its catalogue two models developed by competitors. The first, DeepSeek V4.1 Flash, is presented as “ideal for development tasks that require a large context window and for agentic tasks.” The second, GLM 5.3 Flash, is described as a “more compact” version, faster and less costly than Z.ai’s flagship model.
The Responses API accepts, according to Mistral AI, “text, images or files.” A developer can, for example, ask a multimodal model to summarise a document or describe an image. The model can also call application functions, to query a database or trigger an action, and stream its response as it is generated. The documentation also mentions structured outputs in JSON format, a data format readable by programs. The interface is suitable both for a new project and for an integration already designed for OpenAI’s format.
The API is also “stateless,” meaning it retains no record of the exchanges. It is therefore up to the application to manage conversation memory. “Your application keeps the conversation history and passes it in the input of each request, including the output elements of previous responses,” the documentation states.
An open preview for everyone
The public preview is accessible to all developers with a Mistral API key, starting now. Mistral AI notes, however, that its operation may evolve before the final version. Requests are processed on servers located in the European Union and the United States. The list of features available during this phase is listed in the company’s documentation, under the compatibility section.
