Mistral AI is an AI lab founded in Paris in 2023. It builds large language models and sells access to them three ways: a consumer chat app called Le Chat, a pay-as-you-go API through La Plateforme, and open-weight model downloads you can run on your own hardware. That third option is the thing that makes Mistral different from OpenAI, Anthropic and Google, and it is why this course spends so much time on cost.
Under the hood a Mistral model is a standard large language model. It reads your prompt as tokens (roughly three-quarters of a word each), predicts the next token, appends it, and repeats until it decides to stop. That loop is the whole product. Everything you pay for, whether it is a Le Chat plan or an API invoice, is a wrapper around that loop, so understanding tokens is the fastest way to understand the bill.
The three ways to use Mistral
- Le Chat (chat.mistral.ai) is the ChatGPT-style app: chat, web search, document upload, image generation and code. There is a free tier and paid tiers.
- The API (console.mistral.ai) bills per million tokens. The request shape is OpenAI-compatible, so most existing code ports with a URL and a key change.
- Open weights: many Mistral models (the Small, Ministral, Nemo, Codestral and Devstral lines, among others) are published under the Apache 2.0 licence. You can download them, run them locally, and pay nothing per token.
Why people buy Mistral
The first reason is price. Mistral's mid-tier and small models are consistently among the cheapest hosted options for their quality band, and the price ladder in the next lesson shows that live. The second reason is Europe. Mistral is headquartered in Paris, offers EU hosting, and is a common answer when a procurement or legal team asks for a GDPR-friendly vendor with EU data residency. If your data cannot leave the EU, Mistral is often the shortest path that still gets you a frontier-adjacent model.
The third reason is control. Because so much of the range is open-weight, you can start on the API, prove the workload, then move the same model onto your own servers if volume makes that cheaper. No other major lab lets you do that with its mainstream models. That exit route is worth real money in any negotiation.
Honest limits
- Mistral's largest models usually trail the top OpenAI, Anthropic and Google models on the hardest reasoning and coding tasks. Test on your own work before assuming parity.
- Le Chat's consumer features (search, canvas-style editing, agents) move fast and lag the US apps in some areas. Do not pick it for features you have not seen working.
- Self-hosting is only free if you already own the hardware and the people to run it. A rented GPU can cost more than the API at low volume.
- Model naming is confusing: version numbers and date stamps (2512, 2603) appear in the same list. Lesson 2 untangles it.
Knowledge check
Which of these is unique to Mistral among the big frontier labs?