Mistral Drops a Responses API So Developers Can Ditch OpenAI Instantly

Mistral has released its Responses API preview, giving developers a stateless endpoint that follows OpenAI’s Responses format. It works across the chat models hosted on Mistral Cloud Inference, including Mistral Large 4. The company also added two low-cost Chinese open-weight models for coding, tool use, and other agent workloads.

An OpenAI-shaped route to Mistral

The endpoint supports text generation, image and document questions, structured output, function calls, and streaming. Its familiar request and response structure can reduce the adapter code required to move an OpenAI-based agent to Mistral or route workloads across providers. Mistral documents the supported fields in its API documentation.

Mistral’s endpoint is stateless, so the service does not retain conversation history between requests. Applications must store prior messages and output items, then submit the relevant history with each new call. Longer conversations therefore increase input-token usage unless the application summarizes or trims earlier turns.

Each response contains an output array whose message, reasoning, and function-call items remain separate. Agent frameworks can inspect those typed items directly instead of extracting tool calls from a flattened assistant message.

Tools, files, schemas, and streams

The public preview includes the main capabilities required by agent frameworks:

• Text input and output as strings or message arrays.
• Image input by URL or base64 on vision-capable models.
• File input by HTTPS URL, base64, or uploaded file_id.
• Structured output through JSON mode or JSON Schema using text.format.
• Streaming events for text deltas and function arguments.
• Client-executed function calling with namespace support.
• Model-dependent reasoning control through reasoning.effort.
• Preflight token counting through POST /v1/responses/input_tokens.

The Python SDK authenticates with a Mistral API key and exposes the endpoint through client.responses.create:

import os

from mistralai.client import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

response = client.responses.create(
    model="mistral-large-4",
    input="Write a one-sentence greeting.",
    reasoning={"effort": "high"},
    max_output_tokens=4096,
)

The reasoning setting depends on the selected model and should be omitted when unsupported. Mistral provides Python and TypeScript SDKs, while Cloud Inference serves the endpoint from EU and US regions.

State stays in the application

A client-managed conversation or tool loop follows four steps:

1. Store the user messages and every item returned in response.output.
2. Inspect the output for a function_call item and execute the named function in application code.
3. Append a function_call_output item containing the result and matching call_id.
4. Send the updated history in the next request and repeat until the model returns a final message.

Payload compatibility reduces migration work. Client libraries, authentication, model identifiers, supported parameters, rate limits, and error handling still require testing, especially for applications currently tied to OpenAI-specific SDK behavior.

Two Flash models join the catalog

Mistral Cloud Inference now hosts DeepSeek V4.1 Flash and Z.ai’s GLM 5.3 Flash. Both models publish weights under the MIT license and target repeated coding, planning, and tool-use steps where latency and token costs can compound quickly.

DeepSeek V4.1 Flash processes text and images natively. DeepSeek says it trained the model from scratch on a 45-trillion-token multimodal corpus, trained sparse attention at a 64K sequence length, and extended the context window to 1 million tokens during later training.

Z.ai says GLM 5.3 Flash uses roughly one-third of the attention compute and about one-quarter of the KV-cache capacity required by GLM-5.3. Those vendor-reported figures suggest lower serving overhead for workloads that launch several subagents or tool calls in parallel, though teams should validate throughput and output quality on representative traces.

DeepSeek lists peak pricing on its own API at about $0.30 per million input tokens and $1.20 per million output tokens. Those figures do not establish Mistral Cloud pricing, which developers should check separately. The published download sizes also describe model artifacts; deployment memory varies with precision, quantization, batching, context length, and inference framework.

Where the combination helps

Mistral is pairing a familiar API shape with a catalog that spans its own models and third-party open weights. A team can evaluate Mistral Large 4, DeepSeek V4.1 Flash, and GLM 5.3 Flash through one hosted platform while preserving much of an existing Responses-style agent architecture.

EU and US deployment regions may simplify infrastructure planning for teams with geographic processing requirements. Regional availability alone does not settle compliance obligations, so buyers still need to verify data retention, storage, subprocessors, logging, and contractual terms.

Checks before production

• Run representative traces: Include long conversations, structured output, failed tool calls, images, files, and parallel function calls.
• Preserve typed output: Store all required output items and maintain the correct call_id for each tool result.
• Measure full-loop performance: Compare latency, token usage, schema adherence, tool-call success, and final-answer quality.
• Test history growth: Use the token-counting endpoint to determine when conversations need trimming or summarization.
• Audit integration differences: Confirm authentication, model names, supported parameters, streaming events, retries, and error formats.
• Plan for preview changes: Pin SDK versions, isolate provider-specific code, and avoid placing the preview endpoint directly into critical workloads without fallback handling.
• Verify commercial terms: Check Mistral’s current pricing, regional availability, rate limits, and data-processing terms rather than relying on another provider’s published rates.