Neural Brew

Dev & code

Mistral launches a Responses API compatible with OpenAI's format and serves DeepSeek V4.1 Flash and GLM 5.3 Flash

Mistral AI is launching a public preview of a Responses API that adopts OpenAI's Responses format for building agents, and adding two fast, low-cost Chinese models to its inference offering: DeepSeek V4.1 Flash (preview) and Z.ai's GLM 5.3 Flash (generally available).

AI-generated portrait of TomoAI-generated portrait of Lior

By and

Published · 4 min read

Abstract generated illustration for the Dev & code section

Mistral AI announced this Wednesday, October 28, at 3:01 PM Paris time, the opening in public preview of its Responses API, presented as “designed to build agents on Mistral.” In the same thread, the company is adding two fast, low-cost models to its inference offering: DeepSeek V4.1 Flash, in preview, and GLM 5.3 Flash from Z.ai, now generally available.

OpenAI’s Responses format, at Mistral

Until now, Mistral models were mostly queried through a Chat Completions-style API, which returns a message. The Responses API instead returns a list of separate items: the model’s response, its reasoning, and its tool calls. For an agent application, each step becomes easier to handle in isolation, which is what Mistral is highlighting.

The documentation is explicit: the API lets you use Mistral models “with OpenAI’s Responses format,” and targets both new projects and existing Responses integrations to migrate to Mistral. According to the API reference, the endpoint is POST /v1/responses, with a responses.create method in the Mistral SDK. Mistral states on X that it works with “all chat models hosted by Mistral, including Large 4.” Mistral’s guide to migrating from OpenAI also points to this API for applications already calling client.responses.create().

During the preview, the following are supported: text, images, and files as input (URL, base64 data, or previously uploaded file), structured outputs (JSON mode and JSON Schema), streaming, function calling, and reasoning adjustable via reasoning.effort, depending on the model. A dedicated endpoint counts input tokens before sending a request. The documentation examples use Mistral Large 4.

A stateless API

An important difference from what is often associated with the Responses format: Mistral’s version is stateless. The server does not keep the conversation; the application must return the full history with each request, including output items from previous responses. Functions are executed client-side: the model requests a call, the application executes it and returns the result with the call ID. Mistral specifies that the API is served in the European Union and the United States on its Mistral AI Cloud Inference offering.

Two low-priced Chinese models

Mistral’s pricing grid, checked on October 28, shows:

  • DeepSeek V4.1 Flash 1017 (deepseek-v4-1-flash-1017), in public preview, one-million-token context: $0.30 per million input tokens, $0.007 for cached input, $1.20 for output. Mistral recommends it for long-context coding and agent tasks.
  • GLM 5.3 Flash from Z.ai (zai-glm-5-3-flash), generally available, one-million-token context: $0.15 input, $0.03 cache, $0.50 output. According to its model card, it is a 320-billion-parameter mixture-of-experts multimodal model, with 18 billion active, served “without modification by Mistral.” The developer targets it at sub-agents, tool calls, and coding.

For comparison, Mistral Small 4 is listed at $0.15 input and $0.60 output, and Mistral Large 4 at $1.36 and $4.18.

What remains to be clarified

Mistral announces compatibility with all its hosted chat models, which should include DeepSeek V4.1 Flash and GLM 5.3 Flash, but the documentation does not publish a detailed per-model table and we have not tested it. The accepted reasoning levels, meanwhile, vary by model. The preview status also means the scope may evolve; Mistral invites users to report issues.

Our take

By adopting OpenAI’s Responses format, now a reference for agents, Mistral lowers the cost of switching providers: an application already written for this format can, in principle, be pointed at Mistral without a complete rewrite. The choice of a stateless API leaves memory management to the application, which is simpler to host in Europe but requires more work client-side. The arrival of models from DeepSeek and Z.ai, served by Mistral in its own regions, shows the company is also positioning itself as a host for third-party open models, competing with American inference platforms. What remains to be verified is the actual compatibility with existing integrations, which only practice will tell.

Sources

  1. Mistral AI (X) — announcement of the Responses API and the two new served models Primary source
  2. Mistral Docs — Responses API (public preview) Primary source
  3. Mistral Docs — Responses API reference (endpoints) Primary source
  4. Mistral Docs — migration guides Primary source
  5. Mistral Docs — DeepSeek V4.1 Flash 1017 model card Primary source
  6. Mistral Docs — models overview Primary source
  7. Mistral Docs — API pricing Primary source
  8. Mistral Docs — Z.ai GLM 5.3 Flash model card Primary source

A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.