The complete guide to Mistral AI

Mistral AI is a French company that builds large language models and releases many of them as open weights. It offers them through the Vibe assistant (formerly Le Chat), an API and private deployments. Founded in Paris in 2023 by former Google DeepMind and Meta researchers, it has become one of Europe’s best-known AI companies.

What is Mistral AI?

Mistral AI develops large language models for businesses, developers and consumers, and takes its name from the cold wind of southern France. Its open-weight strategy sets it apart: most of its models can be downloaded and run on your own servers. That addresses concerns such as data sovereignty, customization and vendor lock-in, which slow AI adoption in regulated industries.

Around those models Mistral has built a full stack: Vibe for work and coding, Mistral Studio for developers, Forge for training custom models, plus Mistral Compute, its own European infrastructure. Mistral says it operates in 20 countries and supports more than 125 global enterprises, including Airbus, ASML and HSBC.

Mistral AI models and pricing (September 2026)

Mistral’s lineup changed substantially in 2026, when Mistral Small 4 and Mistral Medium 3.5 absorbed the separate Magistral, Pixtral and Devstral models. These are the main current text models, with list prices in US dollars from Mistral’s API pricing page, checked on 29 and 30 September 2026:

Model	What it is	License	Input per 1M tokens	Output per 1M tokens
Mistral Medium 3.5	Flagship for agentic work, reasoning and coding; text and images; 256k context	Open weights, modified MIT with a revenue limit	$1.50	$7.50
Mistral Large 3	General-purpose multimodal model, 675B parameters (41B active)	Apache 2.0	$0.50	$1.50
Mistral Small 4	Fast model that combines instruct, reasoning and coding, 119B parameters (6.5B active)	Apache 2.0	$0.15	$0.60
Ministral 3 14B	Small model for edge devices and high volumes	Apache 2.0	$0.20	$0.20
Ministral 3 8B	Smaller edge model	Apache 2.0	$0.15	$0.15
Ministral 3 3B	Smallest edge model	Apache 2.0	$0.10	$0.10
Codestral 25.08	Code completion and fill-in-the-middle	Commercial, no public weights	$0.30	$0.90
GLM 5.3 (Z.ai)	Third-party open model hosted by Mistral, 1M context	Third-party open weights	$1.40	$4.40

Documents are priced per page instead of per token: OCR 4.1 costs $4 per 1,000 pages, and Document AI, which adds structured annotations, costs $5 per 1,000 pages.

Mistral Large 3 is cheaper than the newer Medium 3.5, which Mistral recommends for most tasks, including coding. In production, pin a fixed General Availability version such as mistral-medium-3-5 rather than a -latest alias, which switches to newer models automatically. Preview and Labs models can change without notice, and even a pinned model can vary its answers.

Getting started with Mistral AI

Most people start in Vibe (chat.mistral.ai), which you can use with a free account on the web, on mobile, in the terminal and in VS Code. Developers who want to build Mistral into their software start with the API.

API setup and authentication

You first create an account at console.mistral.ai. New accounts start in Mistral Studio’s free mode, with low rate limits and no credit card required. Create an API key with a descriptive name and copy it straight away, because it is shown only once. Anyone who has the key can make calls billed to your account, so store it in an environment variable or a secrets manager.

Making your first API call

You install the official SDK with pip install mistralai. Version 2 moved the client import to mistralai.client, so older examples with from mistralai import Mistral need updating. Here is a minimal example in Python:

import os

from mistralai.client import Mistral

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

response = client.chat.complete(

    model="mistral-medium-3-5",

    messages=[

        {"role": "user", "content": "Explain quantum computing in simple terms"}

    ],

)

print(response.choices[0].message.content)

Platform options and deployment models

Mistral offers a wide range of deployment options, and it says the majority of its customers already run its models in their own data centers and cloud environments. The options are:

• Cloud API: you pay per token through Mistral Studio, and regional endpoints let you choose whether inference runs in Europe or the US.
• Private cloud deployment: Mistral models are available through Azure AI, Amazon Bedrock, Google Cloud Vertex AI, Snowflake Cortex, IBM watsonx and Outscale, and since July 2026 Mistral Medium 3.5 is also in Microsoft Foundry and Copilot Studio.
• On-premise installation: the open-weight models run on your own hardware with vLLM, TensorRT-LLM or TGI, from a single GPU for the smallest models to multi-GPU clusters for the largest. Enterprise customers can also run Vibe itself on-premises.
• Hybrid deployment: you use the cloud API for development and non-sensitive work, and self-hosted models for production workloads that require maximum data sovereignty.

Benefits and advantages of Mistral AI

Cost efficiency and performance optimization

Compared with the OpenAI and Anthropic flagships, Mistral is much cheaper. Mistral Large 3 costs $0.50 per million input tokens and $1.50 per million output tokens, against $10 and $50 for both OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1. Those flagships target the hardest tasks, but for high-volume work such as classification, extraction and summarization, a Mistral model often does the job at a fraction of the cost. Among the low-cost models the gap is much smaller, as the comparison table below shows. Batch processing halves the price and cached input tokens cost up to 90 percent less. Our guide to LLM cost optimization explains both techniques.

Vibe Pro costs $14.99 a month and Team $24.99 per user a month, both excluding tax, while Enterprise is priced on request.

Open source flexibility and customization

Most current Mistral models are open-weight, but not all under the same license. Mistral Large 3, Small 4 and Ministral 3 use Apache 2.0, which allows commercial use and self-hosting without paying Mistral. Mistral Medium 3.5 uses a modified MIT license that does not apply to companies, or employees of companies, with global consolidated revenue above $20 million in the previous month. They need a commercial license or Mistral’s hosted services.

Codestral 25.08 and OCR 4.1 have no public weights, so you use them through the API or a cloud partner. Mistral does offer self-managed deployment of OCR 4 to enterprise customers, but that is a commercial arrangement, not a download of open weights. Mistral Large 2 was not open source either: it used the non-commercial Mistral Research License and has since been deprecated. Its successor, Large 3, is open-weight and uses a Mixture of Experts architecture, so only a fraction of its parameters is active per token.

Open weights let you fine-tune a model on your own data and run it without any connection to Mistral. For deeper customization, Mistral launched Forge in March 2026. It lets enterprises train models on their proprietary knowledge. The overview of the best open-source LLMs compares Mistral with other open model families.

Data sovereignty and security

Mistral’s European base is its strongest argument for regulated industries, but storage and processing are separate questions. According to Mistral’s help center, stored data is hosted in the European Union by default, and Mistral is a French company without a US parent. For inference, the standard global endpoint makes no commitment on location. The EU endpoint keeps processing in data centers in EU and EFTA countries for a 10 percent surcharge, but it does not support Agents, Batch or the Files API, and not every model is available there. Some features can also send data temporarily to subprocessors outside the EU under standard contractual clauses, and Enterprise customers can switch some of them off at organization level. For full control, you run the open-weight models on your own infrastructure.

Check the data settings before you use real data. In Vibe, inputs and outputs are used for training by default on every plan except Enterprise, and the free API mode may use them too. Both can be switched off in the admin panel, with separate settings for Vibe and the API. Feedback you give with the thumbs buttons can still be used, and experimental Labs models do not support the opt-out. Mistral reports SOC 2 Type II and ISO 27001/27701 compliance and offers zero data retention for stateless paid API calls on request. For US healthcare data under HIPAA, confirm Business Associate Agreement coverage with Mistral or your cloud provider first, or host the models yourself.

Multilingual and multimodal capabilities

According to their model cards, Mistral’s current models support dozens of languages, including English, French, German, Spanish, Italian, Dutch, Chinese, Japanese and Arabic. Medium 3.5, Large 3 and Small 4 also accept images. OCR 4.1 handles documents, and the Voxtral models cover transcription and speech.

Limitations and potential downsides

Technical complexity and resource requirements for Mistral AI

Open weights also bring extra work. Self-hosting means you run inference, monitoring, updates and security yourself, which requires AI engineering skills many organizations do not have in-house. Hardware is the second hurdle: the Ministral models run on one GPU, but Mistral Large 3 has 675 billion parameters and needs a multi-GPU cluster. For modest workloads the API is usually cheaper, as the guide to self-hosted LLMs explains.

Ecosystem maturity and support limitations

Mistral has a smaller ecosystem than OpenAI, Microsoft and Google, with fewer third-party integrations and community examples. Vibe does now connect to more than 100 tools. The pace of change is a limitation too: in the first half of 2026 Mistral deprecated Magistral, Devstral, Pixtral and Mistral Large 2.1 and renamed Le Chat. Pinning model versions and following Mistral’s model lifecycle policy keeps those changes under control.

Performance considerations in specific domains

For their price, Mistral’s models are strong, but they do not lead the field. On the Arena text leaderboard of 26 September 2026, which ranks models by user votes in text conversations, the top ten places are all held by proprietary models from Anthropic, Meta and Google. That ranking does not cover every business task, but for the hardest reasoning or coding work a closed frontier model may still perform better, so test on your own data.

Key use cases and real world applications

Enterprise document processing and OCR

One of Mistral’s clearest strengths is OCR 4.1, which extracts text, tables and layout from invoices, contracts, forms and scanned archives for $4 per 1,000 pages, or $2 with the batch API. If you need structured annotations as well, Document AI costs $5 per 1,000 pages. Finance teams use it to automate invoice processing and legal teams to make contracts searchable. In healthcare, a self-hosted open-weight model lets you process patient records without sending them to a third party. Compliance with HIPAA or GDPR still depends on how you deploy and govern it.

Software development and code generation

For coding, Mistral has moved from a separate model to an agent. Vibe’s Code Mode, command-line tool and VS Code extension run on Mistral Medium 3.5 and build features, fix bugs and open pull requests. According to Mistral, the model scores 77.6 percent on SWE-Bench Verified. Codestral 25.08 remains the model for fast code completion inside the editor.

Customer service automation and chatbots

Mistral models power customer service assistants and internal help desks that answer in the user’s language. Together with Mistral, the shipping and logistics group CMA CGM co-developed the agentic AI platform MAIA, Powered by Mistral. The group planned to roll it out progressively from 1 June 2026 to nearly 80,000 employees across CMA CGM, CEVA Logistics and CMA Media.

Data analysis and business intelligence

Organizations use Mistral models to query business data in natural language and generate reports. In Vibe’s Work Mode you can connect a database or upload a spreadsheet and get patterns, anomalies and charts back in the conversation.

Competitor comparison: Mistral vs ChatGPT, Claude and Gemini

“Mistral vs ChatGPT” can mean two things: the API models developers build on, or the assistant apps people use every day. The table covers the API side. It compares selected models from Mistral and three leading US providers, the most capable model and a lower-cost model from each, at list prices in US dollars per million tokens checked on 29 and 30 September 2026.

Feature	Mistral AI	OpenAI (ChatGPT)	Anthropic (Claude)	Google (Gemini)
Most capable API model	Mistral Medium 3.5	GPT-6 Astra	Claude Fable 5.1	Gemini 3.1 Pro (preview)
Price (input / output)	$1.50 / $7.50	$10 / $50	$10 / $50	$2 / $12 (up to 200k tokens)
Lower-cost API model	Mistral Small 4	GPT-6 Luna	Claude Haiku 4.5	Gemini 3.8 Flash
Price (input / output)	$0.15 / $0.60	$0.10 / $0.50	$1 / $5	$0.75 / $3.75 (until the end of 2026)
Context window (most capable model)	256k tokens	1.05M tokens	1M tokens	1M tokens
Open-weight models	Yes, including the most capable model	gpt-oss only	No	Gemma only
Most capable model on your own servers	Yes, with downloadable weights (commercial license needed above $20 million monthly revenue)	No	No	Gemini models on Google-managed hardware on site (Google Distributed Cloud connected, preview), no downloadable weights
EU processing	Stored data in the EU by default; EU inference through the EU endpoint (+10 percent)	EU data residency available	Through cloud partners	Through Google Cloud regions
Headquarters	Paris, France	San Francisco, US	San Francisco, US	Mountain View, US

The main difference is openness. OpenAI, Anthropic and Google keep their most capable models closed, while Mistral releases its flagship as open weights. On price, Mistral’s large models are far cheaper than the OpenAI and Anthropic flagships and somewhat cheaper than Gemini 3.1 Pro, but among the low-cost models GPT-6 Luna is slightly cheaper than Mistral Small 4. On the Arena text leaderboard the US models lead, and compared with Claude, which tops that leaderboard, Mistral trades peak performance for control and European processing options.

The assistant apps are a separate purchase. In a check of the providers’ pricing pages from the Netherlands on 30 September 2026, Vibe Pro cost €14.99 a month excluding VAT (€18.14 including 21 percent VAT) and Claude Pro €18 excluding tax, while ChatGPT Plus and Google AI Pro listed €23 and €21.99 a month. That makes Vibe the cheapest of the four paid plans at monthly billing.

Future outlook and roadmap for Mistral

Model development and capabilities expansion

In 2026, Mistral’s roadmap has focused on consolidation. Mistral Small 4 combines instruct, reasoning and coding in one model, and Mistral Medium 3.5 does the same at flagship level. Mistral also agreed to acquire Koyeb (serverless deployment) in February and Emmi AI (physics AI for engineering) in May, a sign that it is moving into infrastructure and industrial AI.

European AI sovereignty initiatives

Sovereignty is central to Mistral’s strategy, and part of it is already in place. In July 2026 Mistral reached a multibillion-dollar deal under which Microsoft uses Mistral’s European GPU capacity. France’s Ministry of the Armed Forces signed a framework agreement with Mistral in December 2025, and Mistral is a signatory to the EU’s General-Purpose AI Code of Practice under the AI Act. Other parts are still plans: up to 1 GW of computing capacity by 2030 and a data center in Sweden, built with EcoDataCenter, that is scheduled to open in 2027. Our article on sovereign AI explains the wider debate.

Market positioning and competitive strategy

In September 2026 Mistral raised €3 billion at a valuation of more than €21 billion, a year after ASML led its €1.7 billion Series C. The Series D, led by Samsung Electronics with EQT’s Scaleup Europe Fund and PSG Equity as co-leads, is according to Mistral the largest equity round ever raised by a European technology company. Mistral positions itself as the only AI company building the full stack. With open models, its own European compute and its own products, it is a logical first candidate for European organizations that want a sovereign option.

Planning a Mistral AI deployment

Evaluation and proof of concept development

A good evaluation starts with one concrete business problem and a focused proof of concept, tested first with synthetic or redacted data. Before any confidential data goes in, check the agreement that applies to your plan, the retention period, where the data is processed and which connectors are switched on, because switching off training alone is not enough. Compare at least two Mistral models with your current solution, and record a baseline for cost, accuracy and processing time so you can calculate the return.

Technical infrastructure planning

The right deployment model depends on your data. The API or a marketplace such as Azure or Amazon Bedrock is the fastest route. Self-hosting gives maximum control but needs GPU capacity and a plan for model updates, so involve Mistral’s enterprise team early if you plan to self-host at scale.

Team preparation and skill development

Before you scale, your team needs skills and clear governance: which data may go into which tool, who approves new use cases and how you monitor output quality. The EU AI Act requires organizations that use AI to take measures to support the development of their staff’s AI literacy. According to the European Commission, since the Digital Omnibus the Act no longer prescribes a specific level.

Frequently asked questions

Want to learn more about Mistral or see it in action? A live Mistral demo shows what the models and Vibe can do on tasks from your own work. In a Mistral workshop, your team learns to use Mistral in practice. If you want to build and run your own solution, our team can help with Mistral development and implementation.