Mistral AI is a French artificial intelligence lab headquartered in Paris, and you can use it three ways today: through Vibe, its agent for work and code (called Le Chat until May 2026); by downloading its open-weight models, led by Mistral Large 4; or through its pay-as-you-go API, where Large 4 costs $0.70 per million input tokens. I can build models and APIs, European ones like Mistral’s included, into business agents and automations, and teach your team how to maintain them. Here’s what each piece is, what it costs, what “data in the EU” actually means and when I wouldn’t recommend it.
What Mistral AI is today (and which name means what)
Mistral AI was founded in April 2023. Its co-founders are Arthur Mensch (CEO), Guillaume Lample (Chief Science Officer) and Timothée Lacroix (CTO). Its data processing agreement describes it as “a French limited joint-stock corporation, incorporated in Paris,” with registered offices at 15 rue des Halles. Its mission, per its company page: “make frontier AI open to all, and together solve the world’s hardest problems.”
On September 8, 2026, it announced a €3 billion Series D at a post-money valuation of more than €21 billion, led by Samsung Electronics, with Scaleup Europe Fund (managed by EQT) and PSG Equity as co-leads. Mistral calls it “the largest equity fundraising round ever completed by a European technology company.” A year earlier, in September 2025, it closed a €1.7 billion Series C led by ASML. The company says it operates in 20 countries and serves more than 125 large enterprises, including Airbus, ASML and HSBC, and since September 2026 its models have powered Firefox Smart Window (beta) in France and North America.
The confusion on Google comes from the name changes. Here’s the map:
| Name | What it is | Where |
|---|---|---|
| Vibe | Mistral’s agent for work and code. It merges what used to be Le Chat and Mistral Code | chat.mistral.ai |
| Vibe CLI and extensions | The coding agent in the terminal, in VS Code, in JetBrains and in Zed | Inside Vibe |
| Mistral Studio | The enterprise platform: API, agents, workflows, connectors and deployments (formerly La Plateforme) | mistral.ai/products/studio |
| Forge | The service for customizing and training models on your data (March 2026) | mistral.ai/news/forge |
| Mistral Compute | Dedicated compute capacity in Europe for training and inference | mistral.ai/cloud |
| Le Chonk | The official nickname for Mistral Large 4 | This page |
Vibe (formerly Le Chat): what it is and what it costs
On May 28, 2026, Le Chat became Vibe. The product page puts it this way: “Le Chat is now Vibe. All your conversations, settings, and plans carry over.” If you had a Le Chat account, you still have it, along with your conversations and your plan.
Vibe combines two things in one product: Vibe for work (“formerly Le Chat”), an agent for long-running tasks that does research, catches you up on your inbox, drafts documents, spreadsheets and slides, and runs recurring tasks on a schedule; and Vibe for code, a coding agent that works in the terminal (Vibe CLI), in your editor (VS Code, JetBrains or Zed) or on the web, with remote agents that carry work through to a pull request. It connects to more than 100 tools, such as email, calendar, Slack, GitHub and Jira, and it’s compatible with MCP. Voice input runs on Voxtral, Mistral’s speech models.
Official plans from mistral.ai/pricing, checked on November 16, 2026:
| Plan | Price | What you get |
|---|---|---|
| Free | Free | “Limited messages and web searches,” limited coding sessions, basic image generation, 100+ connectors and access to Mistral Studio |
| Pro | $14.99/mo | “All-day coding” in the terminal, the editor or on the web; spreadsheets and slides; $25.5 a month in API credits; chat and email support |
| Pro for students | $5.99/mo | Up to 12 months, for verified students at accredited higher-education institutions who haven’t used Vibe or Le Chat before; $15 a month in API credits |
| Team | $24.99 per user/mo | Higher limits, up to 30 GB of storage per user, domain verification and data export |
| Enterprise | Custom | Custom models, agents and workflows, audit logs, SAML SSO, white label and private deployments |
About currency and taxes. The pricing page has a currency selector (USD or EUR), and the figure is the same in both: $14.99 or €14.99 for Pro, and $24.99 or €24.99 per user for Team. These prices exclude taxes: the page lets you toggle between “Excluding taxes” and “Including taxes,” and the student plan notes the price is “not inclusive of taxes.” I haven’t checked the final amount with sales tax or VAT for any specific country, so look at it before you pay.
Limits. The site doesn’t say how many messages the Free plan includes, only that they’re limited. Its FAQ gives one example of the gap: Pro gets “up to 40x more image generations” than Free.
Your conversations and model training. In the plan comparison table, the “Model training” row shows “Opt-out” on Pro, Team and Enterprise: you can switch it off, but you have to do it yourself. On Free it only shows a checkmark, without the “Opt-out” label, and the page doesn’t say whether you can switch it off on that plan. On Enterprise, Mistral offers to deploy Vibe on your own servers, in a private cloud or on Mistral Cloud “with full data residency.”
Mistral’s current models
This is the lineup on the official models list today. The license is the one Mistral publishes for each model; “Premier” is the label Mistral gives its commercial models.
| Model | What it is | License | Context | Status |
|---|---|---|---|---|
Mistral Large 4 (mistral-large-4) | The flagship: a multimodal mixture-of-experts (MoE) model with 1 trillion parameters, 52 billion active, and three reasoning modes (low, medium and high) | Mistral Ecosystem Open-Weight License | 1 million tokens | Preview on October 6, 2026; open weights on October 29 |
Mistral Small 4 (mistral-small-2603) | A 119-billion-parameter MoE (6.5 billion active) that combines direct answers, reasoning and coding in one multimodal model | Apache 2.0 | 256,000 tokens | Current (March 2026) |
| Ministral 3 (3B, 8B and 14B) | Compact models with vision | Apache 2.0 | 256,000 tokens | Current (December 2025) |
| Codestral | A model specialized in code completion | Premier (commercial) | 128,000 tokens | Current (version 25.08) |
| Mistral Medium 3.5 | A dense 128-billion-parameter model | Modified MIT | 256,000 tokens | Deprecated since October 31, 2026; retires April 30, 2027. Replacement: Large 4 |
| Mistral Large 3 | The previous flagship | — | — | Deprecated since October 31, 2026; retires April 30, 2027. Replacement: Large 4 |
Large 4, in Mistral’s own words. The October 6 announcement describes it as “a 1 trillion-parameter natively multimodal model with 52 billion active parameters,” trained from scratch “on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe.” According to the same post, a significant share of its training data was multilingual, spanning more than 160 languages, “including every official language of the European Union.” The weights came out on October 29 on Hugging Face, and Mistral says you can deploy them on your own infrastructure with vLLM or SGLang.
Beyond text. The API includes document models (OCR 4.1), speech models (the Voxtral 3 family, from October 2026, with transcription and text-to-speech, in public preview), embedding models (Mistral Embed and Codestral Embed) and moderation. Mistral also releases open-weight time series models (Timestral) for running on your own infrastructure. And something new: Mistral now serves third-party models through its API, such as Z.ai GLM 5.3 and DeepSeek V4.1 Flash.
What’s gone. Mistral 7B and Mixtral, the models that put Mistral on the map, no longer appear in the current list. Devstral 2 was retired on July 31, 2026, and the Magistral series has been retired too.

◆ Licenses: Large 4 isn’t Apache 2.0, even if a Mistral page says so. The open-weights announcement and the official models list publish it under the “Mistral Ecosystem Open-Weight License,” a modified MIT license. According to Mistral, the model is “free to use, modify, fine-tune, and commercialize, with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially.” In other words: if your business is offering the model as a service to third parties, you need a separate commercial license. Meanwhile, the summary cards on both Large 4 announcements (October 6 and 29) still show “Apache 2.0,” and the pricing FAQ says open-weight models (“e.g., Mistral 7B”) are Apache 2.0 “for research/individual use,” a line that doesn’t match each model’s actual license. Small 4 and Ministral 3 really are Apache 2.0. Before you build a product on one, read that specific model’s LICENSE file.
Mistral API pricing today
Official prices from the API pricing table, checked on November 16, 2026. These are prices in dollars per million tokens, standard tier.
| Model | Input | Cached input | Output |
|---|---|---|---|
| Mistral Large 4 | $0.70 | $0.07 | $2.10 |
| Mistral Small 4 | $0.15 | $0.015 | $0.60 |
| Ministral 3 14B | $0.20 | $0.02 | $0.20 |
| Ministral 3 8B | $0.15 | $0.015 | $0.15 |
| Ministral 3 3B | $0.10 | $0.01 | $0.10 |
| Codestral | $0.30 | $0.03 | $0.90 |
| Mistral Embed | $0.10 | $0.01 | — |
| Codestral Embed | $0.15 | $0.015 | — |
Other prices from the same table: OCR 4.1 at $4 per 1,000 pages; Voxtral Mini Transcribe 3 at $0.0035 per minute; Voxtral TTS 3 at $13 per million characters; Voice Agents at $0.04 per minute with Small 4 and $0.08 with larger models; and Mistral Moderation 2, free.
What you should know before you pay:
- Batch and caching. According to the pricing FAQ, batch processing “reduces the price by 50%” and cached input cuts input cost “by up to 90%” for repeated prompts.
- The EU endpoint costs 10% more. Regional inference is billed at 1.1× the standard rate (more on that below).
- The launch deal is over. Large 4 launched at 50% off for two weeks starting October 6; today it’s $0.70 and $2.10.
- The pricing FAQ is out of date. It still uses the example “Mistral Large costs $0.5 /M tokens in and $1.5 /M tokens out,” and it recommends “For most tasks and coding: Mistral Medium,” even though Medium 3.5 is already deprecated. The docs pricing table lists $0.70 and $2.10 for Large 4, matching its announcement: budget with that figure.
- Third-party models. On Mistral’s API, DeepSeek V4.1 Flash costs $0.30 input and $1.20 output, and Z.ai GLM 5.3 Flash costs $0.15 and $0.50.
◇ Euro prices: not verified. The API pricing table has a dollar/euro selector, but I only read the dollar amounts. I couldn’t confirm the euro figures or how taxes are applied to your account. Check the final amount in your billing console before you budget for it.
Does your data stay in the EU? What the docs say
This is the point that gets oversimplified the most. Mistral being French doesn’t mean every request to its API is processed in Europe. The regional inference docs list three endpoints:
| Endpoint | Where it processes | Surcharge |
|---|---|---|
api.mistral.ai (global, used if you don’t pick a region) | No guaranteed location: “Mistral does not commit to a specific inference location” | No |
api.eu.mistral.ai | “Multiple data centers in EU and EFTA countries” | 10% |
api.us.mistral.ai | “Multiple data centers in the United States” | 10% |

The fine print on the EU endpoint:
- It only serves models hosted in that region, and the docs tell you to check before sending production traffic. I haven’t verified whether Large 4 is available on the EU endpoint today.
- Function calling is the only supported tool. Agents, Batch and the Files API aren’t available on regional endpoints. If your workload depends on an unsupported feature, the docs themselves warn that the regional processing guarantees don’t apply to it.
- The control plane isn’t regional. In the docs’ words: “Regional inference does not make the Mistral control plane regional.” Account configuration, API keys, billing or usage analytics may be handled outside the region you pick.
Other pieces that do exist: Mistral publishes its data processing agreement and runs a Trust Center, although I haven’t read which specific certifications it lists. For Large 4, its announcement also mentions “a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law.” And with the open weights you can run Large 4 or Small 4 on your own infrastructure: your data never leaves it.
The European angle, without the hype. Choosing Mistral doesn’t make you GDPR-compliant on its own. What it gives you are building blocks: an endpoint that processes in the EU, a published data processing agreement, the option to turn off training on your conversations in Vibe’s paid plans, and models you can run on your own servers. If you handle customers’ personal data or work in a regulated industry, get the guarantees in writing and review the case with whoever handles data protection at your company.
How to use Mistral AI: three paths
1. Vibe, in the browser or on your phone. It’s the fastest way to try the models: the Free plan is enough to get a feel for them. If you code, Vibe CLI and the VS Code extension give you the coding agent.
2. The Mistral Studio API. Pay as you go, with the models in the tables above. If you have data that must be processed in the EU, point at the api.eu.mistral.ai endpoint from day one and check which models it serves. If you want to try Mistral alongside other models with a single key, OpenRouter aggregates several vendors, and Mistral’s open-weights announcement says Large 4 is available there too; I haven’t checked at what price or which other Mistral models it serves.
3. Open weights. Large 4, Small 4 and Ministral 3 are downloads from Hugging Face. Mistral also publishes GGUF versions of Ministral 3, the format desktop tools use; my LM Studio review explains how that works, but check the model is in its catalog, because I haven’t verified it. To customize models on your own data, Mistral offers Forge; in Vibe, fine-tuning on your data is only available for Enterprise deployments, through Forge.
Mistral AI alternatives
Entry prices from each vendor’s official pages, in dollars per million tokens, checked on November 16, 2026.
| Alternative | Reference model | Input | Output | Note |
|---|---|---|---|---|
| Mistral AI (baseline) | Mistral Small 4 · Mistral Large 4 | $0.15 · $0.70 | $0.60 · $2.10 | EU endpoint: +10% |
| Qwen (Alibaba Cloud) | qwen3.8-flash · qwen3.8-max | $0.15 · $2.00 | $0.47 · $6.00 | Singapore list prices. In Frankfurt, the default scope isn’t the European one. Qwen review |
| DeepSeek | deepseek-flash (V4.1-Flash) | $0.15 off-peak · $0.30 peak | $0.60 · $1.20 | deepseek-v4-pro: $0.66 and $1.98 off-peak. DeepSeek review |
| OpenAI | gpt-6-luna · gpt-6.1-sol | $0.10 · $2.00 | $0.50 · $10.00 | Standard tier. OpenAI review |
| Anthropic | Claude Haiku 5.5 · Sonnet 5.5 · Opus 5.5 | $0.10 · $2 · $4 | $0.50 · $10 · $20 | Haiku 5.5: price for prompts up to 100,000 tokens. Claude review |
| Meta (Muse, formerly Llama) | Muse Spark Open (open weights) | Not verified | Not verified | llama.com now leads to Meta’s developer site. Llama review |
Mistral, Qwen or DeepSeek for a low-cost agent? On the fast models they tie on input ($0.15), but on output Qwen3.8-Flash charges $0.47, versus $0.60 for Mistral Small 4 and for DeepSeek off-peak. Mistral’s edge isn’t price, it’s origin: it’s the only one of the three headquartered in the EU. Qwen also offers an EU-only scope in Frankfurt, but it isn’t the default. A curious detail: if you use DeepSeek V4.1 Flash through Mistral, you pay $0.30 and $1.20, which is DeepSeek’s peak-hour rate.
And against ChatGPT or Claude? Large 4 ($0.70 and $2.10) costs far less than Claude Sonnet 5.5 or gpt-6.1-sol ($2 and $10), but not less than OpenAI’s and Anthropic’s small models, gpt-6-luna and Haiku 5.5 ($0.10 and $0.50). I won’t give you a verdict on quality: the performance numbers Mistral publishes are its own. What I can tell you is that in the human coding evaluation cited in the open-weights announcement, Large 4 ranks behind Claude Opus 5.5. Test the candidates on your own data.
My verdict as an operator
If you want a European vendor, Mistral AI covers almost everything: an agent for work and code (Vibe), an open-weight flagship (Large 4), a well-priced Apache 2.0 small model (Small 4) and an API that can process in the EU. For agents and automations where the vendor’s origin or control over the model weighs on the decision, it’s one of the first options I’d put on the table.
What I won’t hide from you: the default endpoint doesn’t guarantee Europe, the EU endpoint costs 10% more and leaves out Agents, Batch and Files, Large 4 has a license condition for anyone selling models as a service, and its own pricing page carries stale numbers. My recommendation: try Vibe for free, measure one real workflow with Small 4 or Large 4 on the endpoint that fits your case and, if sensitive data is involved, look at running the open weights on your own infrastructure.
I’ll build it in and teach you how to use it
I can build models and APIs, European ones like Mistral’s included, into business agents and automations, and train the teams who’ll maintain them. The work would come down to picking the model and endpoint that fit your data, costing it out against your real volume and leaving it running without you depending on me.
- If you want an agent built on the right model: AI agent implementation.
- If you want the model wired into your processes: Automations.
- If you want me to teach you and your team to work with open models and APIs: AI Mentoring.
- If you’re not sure where to start: tell me about your situation and I’ll tell you whether Mistral is a fit or whether something else makes more sense.
Mistral AI FAQ
What is Mistral AI?
Mistral AI is a French artificial intelligence lab. You can use it to chat, research, draft documents, code with agents (through Vibe), analyze images and documents, transcribe speech, and build agents and automations on its API. Its flagship model is Mistral Large 4, from October 2026, with open weights.
Is Mistral AI free?
Partly. Vibe has a Free plan with limited messages and searches, and the open-weight models (like Small 4, under Apache 2.0) are free to download. Vibe Pro costs $14.99 a month before taxes (the same figure in euros), and the API is pay-as-you-go.
What happened to Le Chat?
On May 28, 2026, it became Vibe. According to Mistral, “All your conversations, settings, and plans carry over.” Vibe also adds the coding agent that used to be called Mistral Code.
Who runs Mistral AI?
Arthur Mensch is co-founder and CEO. He founded the company in 2023 with Guillaume Lample (Chief Science Officer) and Timothée Lacroix (CTO).
Who owns Mistral AI?
It’s funded through investment rounds; when it closed its Series C, the company said the round “reaffirms the company’s independence.” The latest, a €3 billion Series D in September 2026, was led by Samsung Electronics; the 2025 Series C was led by ASML. I haven’t verified how its shareholding is split today.
Where is Mistral AI from?
France. It’s a French company headquartered in Paris. In September 2026 it also opened a hub in Munich.
What companies use Mistral AI?
Mistral names Airbus, ASML and HSBC among the more than 125 large enterprises it says it serves, and since September 2026 its models have powered Firefox Smart Window (beta) in France and North America. Those are the company’s own statements.
How much does the Mistral API cost?
Per million tokens, in dollars: Mistral Large 4, $0.70 input and $2.10 output; Mistral Small 4, $0.15 and $0.60; Codestral, $0.30 and $0.90. Batch processing costs half, and the EU endpoint costs 10% more. Heads-up: the FAQ on the pricing page still lists $0.50 and $1.50 for “Mistral Large”; the docs pricing table lists $0.70 and $2.10.
Does Mistral process my data in the EU?
Only if you use the api.eu.mistral.ai endpoint, which processes in data centers in EU and EFTA countries at a 10% surcharge. The global endpoint doesn’t commit to any location. The EU endpoint also doesn’t offer Agents, Batch or the Files API, and account data, keys or billing may be handled outside the region.
Is Mistral Large 4 open source?
It has open weights, but it isn’t Apache 2.0. It uses the “Mistral Ecosystem Open-Weight License,” a modified MIT license: you can use, modify, fine-tune and commercialize it, unless your business is offering models as a service, in which case you need a separate commercial license. Small 4 and Ministral 3 really are Apache 2.0.
What languages does Mistral AI support?
Mistral says a significant share of Large 4’s training data was multilingual, spanning more than 160 languages, including every official language of the EU. I haven’t measured its quality in any specific language myself, so test it with your own content.
What is better than ChatGPT?
It depends on the task, the volume and your data. I don’t have an independent comparison I can cite to say Mistral is better or worse than ChatGPT. What I can compare is API price and origin: Large 4 costs $0.70 and $2.10 per million tokens, versus $2 and $10 for gpt-6.1-sol, and Mistral is headquartered in the EU.
What do you think of Mistral AI?
As an operator: it covers almost everything you’d ask of a European vendor, with chat and a coding agent, an open-weight flagship and a cheap Apache 2.0 small model. Its weak spots: the default endpoint doesn’t guarantee Europe, the European one has limitations, Large 4 comes with a license condition and the pricing page has stale numbers. I don’t publish user ratings I can’t verify.
Can you help me build Mistral into my business?
Yes. I can build models and APIs like Mistral’s into agents and automations, and train the team who’ll use them. You can see how I approach AI agents or message me directly.
Official sources checked on November 16, 2026: Mistral AI · About Mistral · Series D · Series C · Vibe (formerly Le Chat) · Vibe announcement · Vibe pricing · Mistral Large 4 announcement · Large 4 open weights · Models · API changelog · API pricing · Regional inference · Data processing agreement · Trust Center · Mistral and Mozilla · Mistral Compute · Mistral models on Hugging Face. Alternative pricing: Qwen · DeepSeek · OpenAI · Anthropic · Meta.