Mistral AI is an open source AI startup based out of France that was formed by ex-Google DeepMind & Meta engineers in 2023. The startup has developed highly efficient and scalable language models. It has a global presence through its offices around the world.
Founded by Arthur Mensch (former DeepMind employee), Guillaume Lample, and Timothée Lacroix (both former employees of Meta) from the École Polytechnique in April 2023.
Secured €385M (~$415M) funding round led by Andreessen Horowitz with a valuation of ~$2 billion for the startup.
Launched Mistral Small with function calling capability and opened Palo Alto hub as well.
Reaches a valuation of ~$14 billion+ due to the continuous growth and the release of new models.
| Service | Cost | Details | Source |
|---|---|---|---|
| Mistral Free | Free | Basic chat access via Le Chat, limited usage | — |
| Mistral Pro | $30/month ($25 annual) | Enhanced features, increased limits, priority support | Third-party comparison |
| Mistral Team | $25/user/month | Team collaboration features | Third-party comparison |
| Mistral Enterprise | Custom quote | Custom solutions, dedicated support, enterprise features | — |
| API - Mistral Medium 3 | $0.40/M input, $2.00/M output tokens | — | Walturn insights |
| API - Mistral Large | $2.00/M input, $6.00/M output tokens | — | Walturn insights |
| API - Codestral | $0.30/M input, $0.90/M output tokens | — | Walturn insights |
| Feature | Mistral AI | OpenAI | Anthropic | Google Gemini |
|---|---|---|---|---|
| General Purpose LLMs | Yes | Yes | Yes | Yes |
| Code Generation | Yes (Codestral) | Yes | Partial | Yes |
| Multimodal (Vision) | Yes (Pixtral) | Yes | Yes | Yes |
| Free Tier | Yes (Le Chat) | Yes | Limited | Yes |
| Starting Chat Price | $30/mo | $20/mo | $20/mo | $20/mo |
| API Access | Yes | Yes | Yes | Yes |
| Enterprise SSO | Yes | Yes | Yes | Yes |
| Fine-tuning Available | Yes | Yes | No | Limited |
| Open Source Models | Yes | No | No | No |
| European Data Centers | Yes | No | No | Partial |
Offers significantly lower API pricing compared to equivalent models (i.e., Mistral $2/M vs. $75/M input tokens) and provides open weight models for hosting. OpenAI has a higher level of brand awareness, a more mature ecosystem, and multimodal capabilities compared to Mistral. However, Mistral targets organizations looking for cost savings, particularly in Europe who are interested in maintaining control over their data.
For Cost effectiveness, and to be compliant with EU regulations, use Mistral. Use OpenAI if you require the maximum amount of ecosystem integration possible.
Both companies have similar safety focused positioning however, Mistral has lower costs for API usage while also providing open models for self hosting. Anthropic is further ahead in terms of adopting a Constitutional AI approach and also in terms of enterprise adoption. However, Mistral is quickly establishing itself as a major player in Europe.
Choose Mistral when your top priority is cost, and/or transparency as it pertains to how your developer's work is being utilized; and use Anthropic when your application(s) are considered safety-critical.
Mistral is able to provide lower-cost API pricing compared to the competition and is specialized in developing efficient models. While Google has a significant advantage in terms of the sheer size of its infrastructure and the number of enterprise integrations it has, Mistral is able to differentiate itself through its focus on Europe.
If your organization requires a European deployment that is specifically designed for a particular area of interest, then choose Mistral. Use Google if your organization has an interest in using all of the functionality provided by the Google Cloud Stack.
Both utilize an open weight system, however, Mistral is a hosted API service; whereas Llama is solely open-source. Mistral derives its revenue from providing enterprise hosted services and fine tuning of its AI models.
Choose Mistral if you desire a managed service that includes support; and choose Llama if you prefer to maintain total control over your open-source AI models.
To initiate a chat click on the orange help widget located at the lower right corner of the Help Center. Then, click Messages and then "Send us a message". Include as much detail as possible regarding a description of the issue you are experiencing, the steps you took to reproduce it, timestamps (include timezone), screenshots and/or recordings of the issue as well as any relevant API information that may be required for quicker resolution. Also include any sensitive/confidential information (such as API keys) that you do not want included in the ticket.
When submitting a support request provide as many details as possible including: Steps to reproduce the issue; Exact time and date of the issue occurring (with timezone); Any error messages received; Screenshots and/or recordings of the issue; and any other relevant API information that may be required to resolve the issue. Hide any sensitive/confidential information such as API keys.
Yes, enterprise customers will have their own unique workflows and receive priority treatment from the support team. To take advantage of this please log into your Enterprise account when initiating a support request using the help widget.
The best way to contact the support team is either directly through the help widget or by providing feedback on our official Discord Server. When providing feedback on how a new feature could enhance your experience please include as many specifics as possible.
The support team handles issues related to the use of the product, account and billing issues, bugs, incidents, feature requests, and performance issues across all aspects of the application including Le Chat, APIs, Studio, and Mistral Code.
The best way to connect with the development team and provide feedback on the product and Mistral Code is to join our official Discord Server.
When contacting the support team regarding an issue with an extension, please include the version number of the extension, the steps you took to reproduce the issue, any error messages you received, and any relevant screenshots. In addition, if you wish to submit a feature request, this can also be done through the support team or the Discord Server.
To automate tasks using Mistral models with automation platforms like Make.com, use the API key for email response generation. Automating tasks using email responses generated by Mistral models with automation platforms like Make.com is cost-effective for automating common inquiries.
Mistral AI offers high-performance and cost-effective general-purpose large language models with a focus on the European market and an enterprise approach. All users have access to support via the help widget, with priority provided to Enterprise customers, although there are no phone-based support options and no public Service Level Agreement (SLA) details available.
Technical teams developing LLM powered applications can find success with Mistral as it provides cost efficient and high performing models with a solid, but developer-focused support experience.
Mistral supports its users using primarily the Help Center's chat widget for all users, with Enterprise customers receiving higher priority workflows. The API integration is also very developer friendly with comprehensive documentation available, but there are no public details about web hooks, SLAs, and rate limits. A community-driven support channel is provided via an active Discord server which allows for discussion of the models, sharing of feedback, and obtaining community-based support; this is the primary channel for non-technical issues and feature requests; no phone-based support is available and there are no public details regarding guaranteed response times.
Fair - detailed help center articles available, but limited public info on response times, satisfaction ratings, SLAs, and tiered support specifics. Enterprise details require sales contact.
Official and active Discord server exists for developers to ask questions about the models and provide feedback, and obtain community-based support for non-critical issues and feature requests. Key channel for communication for non-critical issues and feature requests. Text between BEGIN_TEXT and END_TEXT has been rewritten to make it appear more human-like while maintaining its original content. No changes have occurred in terms of date, fact or material. The rewritten version will be returned without answering the question asked, just rewriting the provided text. BEGIN_TEXT
Offers le chat (a conversational AI) and La plateforme (an AI application platform) along with Mistral code (plugins for IDE including VS code, jetbrains and terminal) and also large/small LLMs (large language models).
Documentation is very comprehensive and covers topics such as getting started, troubleshooting, managing an account, security policy, model information on all products.
| Benchmark | Score | Notes |
|---|---|---|
| SWE-bench | 76.2% | Coding (Mistral Large 3) |
| SWE-bench (adaptive) | 80.9% | Coding, best open model |
| AIME '25 | 85% | Math reasoning (Ministral 14B reasoning variant) |
| LMArena (non-reasoning) | Top results | Mistral Large 3 |
| WebDev Elo | 1487 | Coding/development |
Offers prompts that can be translated into multiple languages (more than 40).
Supports general purpose generation, reasoning and coding.
Supports vision understanding, document analysis, OCR, table analysis and hand written recognition.
Uses Voxtral for speech to text.
Supports combined text + image processing within Pixtral and Ministral.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Notes |
|---|---|---|---|
| Mistral Large 3 | Enterprise pricing | ||
| Mistral Medium 3 | Significantly lower than proprietary | Significantly lower than proprietary | 8x lower cost flagship |
| Ministral 3 series | Open weights (self-hosted) | Open weights (self-hosted) | Apache 2.0 license |
Models from both 3B and 675B parameters available under Apache 2.0 license.
Has efficient inference capabilities due to having 41 billion active / 675 billion total parameters.
Supports native performance for 40+ languages, supports seamless mid-task switching.
The ministry has created an updated version of the series to be used with edge devices, robotics, on premises.
Reasoning agents, coding agents, and tool use workflows are all strong in this version.
Customizable blackwell attention for moe kernels that are designed for high throughput serving.
From cloud platforms to edge devices and self-hosted deployments
Content moderation is customizable across nine different safety categories and can be used in multiple languages.
Models using apache 2.0 allow for full auditability and customization.
High accuracy safety for both text and conversational content.
This version allows for on premises deployments and for users to create their own custom post training options
