On 28 October, Mistral AI posted a short thread on X: it is “introducing the Responses API in public preview, made for building agents on Mistral”. If you read that and thought “I have no idea what that means”, this explainer is for you.
First, what is an API?
An API is the way one piece of software talks to another. When an app uses an AI model behind the scenes, it sends a request through the model maker’s API and gets an answer back. You never see it, but every AI feature in an app runs through one. Mistral’s announcement is aimed at the developers who build those apps, not at people chatting with Le Chat.
And what is an “agent”?
A chatbot answers. An agent can also take steps: look something up, run a calculation, call another service, then answer. To do that, the model has to be able to say “I need a tool for this” and wait for the result.
So what does the Responses API do?
It is a new way for apps to talk to Mistral’s models. Every response is split into separate items: the model’s answer, its reasoning and any tool calls. Mistral says that lets an app handle each step more easily than with its Chat Completions interface, the existing style, where the reply comes back as a single message. The documentation says the tool calls are run by the developer’s own code, not by Mistral.
Why does it copy OpenAI?
OpenAI introduced its own Responses API in March 2025, as a replacement for its Assistants API. Mistral’s documentation says its version lets developers “build with Mistral models using the OpenAI Responses format”.
Think of it as a common plug. If an app was built for OpenAI’s format, moving it to Mistral takes less rework. Crypto Briefing calls this the most immediate effect: lower switching costs, and the option to send some tasks to one provider and some to another.
What is the catch?
Mistral’s version is stateless. Picture a help desk that keeps no notes: every time you call, you tell the whole story again. With Mistral, your app keeps the conversation and sends all of it with each request.
That is the main practical difference from OpenAI’s version, which can manage conversation state for you, by chaining requests to an earlier response or through its Conversations API. OpenAI saves response objects for 30 days by default. With Mistral, developers porting agent code have to store and resend that history themselves. Crypto Briefing notes that this gives teams control but also means more engineering work, and that the API is still a preview.
Why does it matter for Europe?
Mistral says the API works with all the chat models it hosts, including its flagship Large 4, and that it is served in the EU and the US on Mistral AI Cloud Inference. For a European company or public body, that means agent software written for OpenAI’s format can now run on models served from the EU.
Anything else?
In the same thread, Mistral added two fast, low-cost models from other developers: DeepSeek V4.1 Flash, in public preview, and GLM 5.3 Flash, now generally available. DeepSeek, the Hangzhou lab, released V4.1-Flash as an open model in September, and Z.ai published GLM-5.3-Flash with open weights in August.
In one sentence
Mistral now speaks the same “language” for building agents as OpenAI, so apps can switch more easily, but with Mistral the app has to remember the conversation itself.
