Mistral launches its Responses API in public preview: OpenAI Responses-compatible, returning the answer, reasoning and tool calls as separate items for building agents

What happened

At 14:01 UTC on Oct 28, Mistral AI's official account announced the Responses API in public preview on X, describing it as "made for building agents on Mistral": each response returns the model's answer, its reasoning and its tool calls as separate items, so apps can handle every step more easily than with Chat Completions. A follow-up post says it is served in the EU and US on Mistral AI Cloud Inference.

Key points from the official docs

• Format: it uses the OpenAI Responses format, for new apps or for bringing an existing Responses integration to Mistral.
• Stateless: the app keeps the conversation history and sends it in input with each request, including output items from earlier responses.
• Output structure: the output array can hold message, reasoning and function_call items.
• Features: text input and output; image input (URL or base64, with a vision-capable model); file input (HTTPS URL, base64 or uploaded file ID); JSON mode and JSON Schema structured output via text.format; streaming events for text and function arguments; client-executed function calling, including namespaces; reasoning.effort and plaintext reasoning, depending on the model.
• Counting: POST /v1/responses/input_tokens counts input tokens before generating.
• Examples: the docs use mistral-large-4 with high reasoning effort and a 4,096-token output budget, with the SDK installed via pip install -U mistralai or npm install @mistralai/mistralai@latest.

Models added at the same time

In the same thread Mistral said it now serves two fast, low-cost models on its cloud: DeepSeek V4.1 Flash in public preview (pitched for long-context coding and agentic tasks) and GLM 5.3 Flash, generally available (pitched for subagents, tool calls and coding tasks).

What it means for developers

OpenAI's Responses format is becoming a shared agent interface across vendors. Teams that already built agent loops on Responses can add Mistral as a fallback or an EU data-residency option at low cost. Keep in mind this is a preview and it is stateless: there are no server-side conversations, so history and tool results are yours to manage.