Mistral AI opened its Responses API to public preview on Wednesday, giving developers running agentic pipelines on OpenAI’s format a path to European-hosted infrastructure that requires changing a base URL and swapping an API key — not rewriting orchestration logic. The endpoint, available at /v1/responses on La Plateforme (api.mistral.ai), works across every chat model Mistral hosts, including the recently released Large 4 flagship. The company announced the preview on October 28, 2026.

The launch is a meaningful shift in the developer calculus for organizations with European data-residency requirements. Until now, the switching cost from an OpenAI-based agent stack to Mistral included rewriting the orchestration layer — the client code that parses model output, routes tool calls, and manages conversation turns. That barrier disappears with the Responses API: any developer who has already built against the OpenAI Responses format can point the same code at Mistral’s servers. The Mistral Responses API documentation details the full endpoint specification.

How the Typed-Item Architecture Differs From Chat Completions

The technical distinction between Mistral’s legacy Chat Completions interface and the new Responses API is not cosmetic. Where Chat Completions returns a single blended message object — the model’s reply, any tool invocations, and reasoning traces merged into one response structure — the Responses API delivers a sequence of separately addressable typed items. A reasoning trace arrives as a ResponseOutputItemReasoning item. A function invocation arrives as a ResponseFunctionToolCall item. The model’s prose answer arrives as a ResponseOutputMessage item. Each element has its own schema and can be inspected, logged, or re-processed independently. NVIDIA’s NeMo Gym engineering documentation explains how the Responses API evolved beyond Chat Completions and why typed-item output is necessary for agentic workloads.

That architecture was designed for the realities of agentic loops: an autonomous agent running a multi-step task needs to extract a tool call cleanly, execute it, inspect the result, and feed the outcome back into the next request — all without string-parsing a blended blob of text. The typed-item model makes that loop deterministic and debuggable in ways that the older format could not match. The New Stack’s coverage of OpenAI’s Open Responses standard traces how the Responses format emerged as the successor to Chat Completions for agentic development.

Mistral’s implementation is explicitly stateless. The company’s servers retain no conversation history between API calls; each request must include whatever prior context the application needs. That design choice trades client-side complexity — the developer must manage the conversation buffer and resend prior output items in the input field of each request — for portability and predictability. An application built on a stateless API carries no stored state that binds it to a particular vendor, making it easier to move between Responses-compatible providers or to run the same pipeline against multiple backends in parallel. The Mistral Responses API documentation covers the stateless request model in full.

Capabilities Confirmed at Launch

The /v1/responses endpoint on La Plateforme supports text, image, and file inputs, so a single request can bundle mixed content rather than requiring separate calls for each modality. Streaming, function calling, structured JSON output (via text.format), and full multi-turn conversation management — handled client-side — are all confirmed at launch. Full capability details are listed in the Mistral Responses API documentation.

Developers who want to control reasoning depth on supported models can pass a reasoning.effort parameter (set to high, medium, or low); the API also exposes a separate endpoint at /v1/responses/input_tokens for pre-request token counting. Official Python and TypeScript SDKs are available. The Mistral Responses API documentation includes SDK installation guides.

Mistral Large 4 features prominently in the documentation examples. The model, described in Mistral’s official Large 4 launch announcement as a 1.05-trillion-parameter mixture-of-experts architecture with approximately 49 to 52 billion parameters activating per forward pass, entered public preview on October 6. That parameter range reflects a discrepancy between Mistral’s launch post (49 billion active) and its model documentation page (52 billion active); both figures originate from Mistral itself. Pricing for Responses API calls follows Mistral’s standard per-token rates by model; Large 4 is listed at $1.36 per million input tokens and $4.18 per million output tokens at standard rates, with preview discounts available.

How Mistral Fits Into the Open Responses Ecosystem

OpenAI introduced its Responses API in March 2025 as a purpose-built successor to Chat Completions for agentic workloads. The New Stack’s coverage of the Open Responses standard explains how the older format was architected around turn-based conversation and could not natively represent the interleaved reasoning, tool invocations, and structured outputs that autonomous agents produce. The Responses API resolved that mismatch by making output a sequence of typed items rather than a single bundled object.

That architecture attracted adoption beyond OpenAI. In January 2026, a vendor-neutral specification called Open Responses launched at the Open Responses specification site, backed by OpenRouter, Hugging Face, LM Studio, vLLM, Ollama, and Vercel, formalizing the request-and-response shape so any provider could implement it consistently. Compliance tests are available as a web app. The intent was to decouple the format from any single company’s proprietary control and give developers a portable standard they could build against regardless of which model provider they used.

Mistral’s adoption expands that ecosystem to include a major independent provider at the one-trillion-parameter reasoning scale. A developer building an agent pipeline against the Open Responses format can now route requests to Mistral’s European infrastructure, to a self-hosted vLLM instance, to an Ollama local deployment, or to OpenAI — using a single client integration, without modifying the orchestration layer for any of those backends. The portability extends beyond the Mistral-versus-OpenAI comparison that most coverage of this launch emphasizes; it reaches across the entire emerging ecosystem of Responses-compatible providers.

Responses API as the Keystone of Mistral’s Expanding Agentic Stack

The Responses API launch sits at the center of a sequence of developer-facing infrastructure moves Mistral has made over the past 18 months. The company released its Mistral Agents API announcement in May 2025, providing the first dedicated interface for running AI agents against Mistral models. In April 2026, Mistral shipped Workflows — a durable orchestration layer, built on Temporal, that handles multi-step processes requiring pauses, external API interactions, approval checkpoints, and resumption across model calls. InfoQ’s coverage of Mistral Workflows details the enterprise deployment architecture. On October 20, the company released Mistral Voxtral 3, a voice pipeline combining speech-to-text, text-to-speech, and a voice-agent SDK priced at $0.04 per minute for the full pipeline; Voxtral Small Transcribe 3 ranked third on the Artificial Analysis speech-to-text leaderboard at launch.

The Responses API slots in as the stateless, portable complement to those offerings. Where Workflows provides server-managed durable orchestration for long-running processes that need Mistral’s infrastructure to hold state, the Responses API gives developers who prefer client-side control — and portability across providers — a clean, standards-compatible way to wire up their own agent loops.

What European Data Residency Now Means for Agent Builders

Mistral AI is headquartered in Paris and processes data in European data centers — Large 4 was trained on NVIDIA Grace Blackwell GPUs in Mistral’s European facilities. For organizations operating under GDPR or enterprise data-sovereignty requirements, that geography has always been a potential advantage over US-headquartered providers. The practical problem was that switching from an OpenAI-based agent stack to Mistral required rewriting the orchestration layer, and that engineering cost was often large enough to make the evaluation difficult to justify.

The Responses API removes that barrier. The remaining differentiation arguments — EU data residency, Mistral’s open-weight model releases, and the company’s independence from US provider consolidation — are now accessible to any developer who can update a base URL. Mistral’s September 2026 Series D funding announcement confirmed a raise of €3 billion (approximately $3.35 billion USD at prevailing October 2026 exchange rates, though rates fluctuate and conversions are approximate) at a valuation of approximately €21 billion (approximately $23.5 billion USD), signaling investor confidence that European-headquartered AI infrastructure has a durable role in enterprise deployments.

Is ‘Does Your AI Provider Know Your Country’s Laws’ Now a Procurement Question?

For enterprise buyers, the Mistral Responses API launch surfaces a question that has grown louder as agentic pipelines handle more sensitive business data: where is this inference happening, and under what legal jurisdiction does the provider operate? Mistral’s European incorporation and data-center geography mean that its infrastructure operates under EU law and is GDPR-native. That is not a guarantee of privacy in every sense, but it is a structurally different legal context from a US-headquartered provider subject to the Cloud Act or a provider with significant data-center exposure in jurisdictions with mandatory state-cooperation laws. The Mistral Responses API documentation details the data-handling model.

The Responses API format’s portability makes that question easier to act on: an organization that decides European infrastructure matters for its agentic workloads no longer needs to commit to a wholesale migration to explore it. The answer is now one endpoint change away.

Frequently Asked Questions

How is Mistral’s Responses API different from Chat Completions?

Chat Completions returns a single, blended message object — the model’s reply, any tool calls, and reasoning traces are merged together in one structure that the client must parse. The Responses API returns a sequence of typed output items: a reasoning trace comes back as a ResponseOutputItemReasoning item, a function call comes back as a ResponseFunctionToolCall item, and the prose answer comes back as a ResponseOutputMessage item. Each has its own schema and can be processed independently. That typed-item model makes agent loops more deterministic and easier to debug, because the client does not need to string-parse blended output to extract tool calls or reasoning steps. NVIDIA’s NeMo Gym engineering note on the Responses API explains the architectural distinction in detail.

Can I use my existing OpenAI Responses API code with Mistral?

Yes, in most cases, with two changes: update the base URL to api.mistral.ai and replace the API key. Mistral implements the same request-and-response JSON schema as OpenAI’s Responses API, so orchestration logic, function schemas, and output parsing code written for OpenAI should work against Mistral’s endpoint without modification. Mistral’s implementation is stateless — the server does not retain conversation history — which matches the behavior of OpenAI’s Responses API when not using server-side state storage. The Mistral Responses API documentation lists the complete endpoint specification and compatibility notes.

What is the Open Responses specification, and why does Mistral’s adoption matter?

Open Responses is a vendor-neutral specification launched in January 2026 at the Open Responses specification site that formalizes the Responses API request-and-response format so any provider can implement it consistently. It was backed at launch by OpenRouter, Hugging Face, LM Studio, vLLM, Ollama, and Vercel. The goal is to give developers a portable standard: build an agent pipeline once against the specification, and it should run against any compliant provider without orchestration changes. Mistral’s adoption expands the ecosystem to include a major independent provider at the one-trillion-parameter scale, which strengthens the case that the format is a genuine industry standard rather than an OpenAI-derived compatibility layer.

Which Mistral models work with the new Responses API?

The Responses API works with all chat models Mistral hosts on La Plateforme. Mistral Large 4 — a 1.05-trillion-parameter mixture-of-experts model in public preview since October 6 — is the primary example in the documentation, but smaller and mid-range Mistral models are also supported. The reasoning.effort parameter and the reasoning output item type are available only on models that support extended reasoning; the Mistral Responses API documentation specifies which models include that capability.