Mistral AI API

Mistral AI is a French lab that sells hosted models and also publishes open weights. On 2026-10-31 its flagship Large 4 costs $0.70 in and $2.10 out per million tokens with a 1M context. Its data terms let it train on content unless you opt out, and its API has zero data retention on request.

Facts at a glance

Prices checked 2026-10-31. For the models in depth, see Mistral models. This entry is about the supplier.

Models and prices

Large 4 is a mixture-of-experts model with about 1 trillion parameters in total and about 49 billion active per token, 128 routed experts with 4 active plus a shared one, and text and image input. It has configurable reasoning effort: low, medium and high, with high as the default. Weights download at about 1.06 TB, or 610 GB as an NVFP4 build. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres, and that it is fluent in 160 or more languages including every official EU language. Those are the card's claims, and we have not tested them.

At $0.70 and $2.10, Large 4 sits close to DeepSeek V4 Pro's off-peak rate of $0.66 and $1.98, and Mistral shows no peak surcharge on its page. Against Claude Sonnet 5.5 at $2 and $10, it is roughly a third of the input price and a fifth of the output price. Output at three times input is the usual pattern, so reasoning effort is a cost control as well as a quality control. Send easy requests at low effort.

Medium 3.5 is leaving. The model docs list Medium 3.5 as deprecated on the API on 31 October 2026, which is today's check date. If you built on it, plan a move to Large 4 and re-test, since the price and behaviour differ.

The licence question

Open weights do not mean unlimited use. Large 4's licence grants the usual MIT rights and then adds a clause. If you or an affiliate operate a Model-as-a-Service business, meaning you give third parties inference or post-training access with meaningful control over inputs, parameters or training data, you need a separate commercial licence before any commercial use. End-user products where the model sits behind specific features are excluded, and so is mere request relaying. Internal use, and anything through Mistral's own products, falls outside the rule. Small 4 is plain Apache-2.0, which is the safest choice if licence risk matters.

If you plan to resell inference, read the licence text itself. We summarise it here from the model card, and we are not lawyers.

Plans and the agent product

Mistral sells an assistant and coding agent under the name Vibe. The pricing page lists:

The pricing page also states a student education plan, for verified students at accredited higher institutions, valid for a maximum of 12 months. Vibe has a web interface, a CLI described as a terminal-native coding agent powered by Devstral, and a plugin for VS Code and JetBrains. The page lists API credits inside some paid plans, so one subscription can cover both chat and some API use.

Rate limits

We did not find a published table of Mistral API rate limits on the pages we read. Mistral's help centre and console show limits per workspace and tier. Until you check, assume you will need to ask for more capacity before a launch, as with any lab.

Data retention and training

The data terms are split by product, and the defaults differ.

The data processing addendum describes Mistral acting as a controller for three purposes: training its models, automated moderation including abuse monitoring on its APIs, and usage statistics that exclude customer data and outputs. It adds that abuse monitoring does not apply when zero data retention has been activated. After the service ends, Mistral deletes or returns personal data under its deletion procedures. It names the retention period as the duration of the agreement.

Three facts follow for a buyer. Opt-out is the default posture to check, not opt-in. Zero data retention exists and ties to abuse monitoring, so ask for it by name. And the thumbs-up button is a data-sharing action, so tell your team. Mistral's help centre says its subprocessors outside the EU go through a security review, which in some cases includes measures such as zero data retention or encryption.

Regions

Mistral emphasises European hosting and EU training. We did not find a region parameter on the API pages we read, so treat the region as something to confirm in writing. If you need data to stay in the EU, the open weights are a second route: run Small 4 or Large 4 on your own infrastructure or on a European GPU host.

Which agents can use it, and how

Mistral's own agent is Vibe. To point Vibe at a self-hosted vLLM server, the Large 4 and Medium 3.5 model cards show installing the mistral-vibe tool, setting the active model in ~/.vibe/config.toml, and giving it the context length, 1,048,576 for Large 4. Mistral says Medium 3.5 replaced Devstral 2 as Vibe's model.

For third-party agents, Mistral's hosted API is OpenAI-style over HTTPS, and agents with a custom-endpoint option can use it. See OpenCode, Cline and Aider. Whether each handles Mistral's tool-call format cleanly is not documented in the places we read.

A worked cost example

Take a job with 10 million input tokens and 2 million output tokens, with no caching. The arithmetic is ours, from the published prices.

If a large share of the input repeats, Mistral's cached input at $0.07 per million turns 8 million cached tokens into $0.56 instead of $5.60. The comparison above is a price comparison only. It says nothing about quality on your task, and the cheapest row is not the winner until it passes your tests. What it does show is that Large 4 is priced as a mid-tier model with a flagship-sized context window, which is an unusual combination.

Serving Mistral yourself

The open weights give you a second way to use Mistral that does not depend on its hosted terms. The model cards show serving through vLLM and note a speculative decoding draft head on the Large 4 checkpoint for faster serving, plus an NVFP4 build for Blackwell GPUs with a smaller footprint. Mistral says the NVFP4 build keeps the same performance, which is its claim and not something we measured.

Large 4 is not a single-GPU model. A download of 610 GB even in the smaller build means a multi-GPU machine, and that is a real cost against the API's per-token pricing. The break-even depends on your volume and your hardware, and we did not calculate one because the inputs are yours. For the hardware side, see GPU requirements for LLMs and vLLM. Teams that need data to stay in the building but cannot run a trillion-parameter model should look at Small 4 first.

Open weights compared

Mistral is one of several labs with open weights. The licence is the thing that differs most.

For application builders who do not resell inference, all four are workable. For anyone offering model access to third parties, Mistral's Large 4 clause is the one to read, and the Apache and MIT models are the simpler choice. See open source LLM for the full list with sizes and licences.

Moving off Medium 3.5

Because the docs list Medium 3.5 as deprecated on the API from 31 October 2026, the practical task for existing users is a migration.

1. Switch the model id to the Large 4 id, mistral-large-4, in a staging environment.
2. Re-run your evaluation set. Large 4 is a different architecture, and the reasoning effort setting can change both quality and cost.
3. Re-check cost. Medium 3.5's per-token price was not on the pages we read, so compare against your own invoice.
4. Check context handling. Large 4's 1M window is four times the 256K of Medium 3.5, so prompts you previously had to trim can now fit. Long prompts still cost money.
5. Update agents. Mistral says Medium 3.5 replaced Devstral 2 in Vibe, so check the model name in ~/.vibe/config.toml if you point Vibe at your own server.

If Mistral's deprecation schedule is a concern, pinning to an open-weight model you host yourself removes it. A file on your disk does not expire.

Alternatives

Who should pick Mistral

A pre-launch checklist

1. Opt out of training in the console and in Vibe, and record the setting.
2. Ask for zero data retention if you handle customer data.
3. Confirm the hosting region and any EU-only endpoint in writing.
4. Move off Medium 3.5 before the deprecation bites.
5. Test reasoning effort levels on your tasks, since price follows output.
6. Read the Large 4 licence if you offer AI to third parties.
7. Check the rate limits in the console before launch.

Our take

Mistral is the credible European option, with real open weights, a clear licence for most application builders, and a price that undercuts the US labs. The documents also show where the work is. Training is on by default unless you opt out, the API terms point to a zero-retention route without explaining it on the pages we read, and we could not find a rate-limit table or a region parameter. A buyer who needs EU processing should get those answers in writing, and a team with strict data rules can fall back on Small 4 weights under Apache-2.0.

We have not run tests for this page. It reflects Mistral's published pages on 2026-10-31. Mistral does not pay for placement here.

Related pages

Mistral models, LLM API prices, open source LLM, private AI for teams, vLLM.