The open-source AI darling just adopted the same data playbook as OpenAI and Google, only the price floor for privacy is higher.
The Summary
- Mistral changed its privacy policy to train on user inputs and outputs by default across free and paid tiers, with only enterprise customers able to opt out
- The shift mirrors OpenAI's approach but comes from a company that positioned itself as the European, privacy-first alternative
- If you're building on Mistral's API below enterprise pricing, your prompts and your users' prompts are now training data
The Signal
Mistral updated its help documentation to make explicit what many AI companies bury in terms of service updates. Free tier users and standard paid subscribers now contribute their input and output data to model training. Only enterprise customers can opt out. The policy applies across Mistral's API and its consumer chat interface, Le Chat.
This is a straight borrow from the OpenAI playbook. ChatGPT Plus subscribers ($20/month) have their data used for training unless they manually opt out in settings. But Mistral's enterprise tier pricing isn't public, and their standard API pricing runs $2-$8 per million tokens depending on model size. The privacy tax is real, and it's not cheap.
"The open weights company just adopted closed data practices."
The timing matters. Mistral raised $640 million in June 2024 at a $6 billion valuation, positioning itself as Europe's answer to American AI dominance. Part of that pitch was GDPR compliance, data sovereignty, and a more transparent approach to model development. Training on user data by default doesn't break GDPR, but it does break the implied social contract with developers who picked Mistral specifically because it wasn't OpenAI.
For builders, this changes the cost calculus of using Mistral:
- If you're prototyping, your ideas are now training data
- If you're running a production app on standard pricing, your users' prompts feed the next model version
- If you want true data isolation, you need enterprise pricing or self-hosted deployment
The self-hosted option is real, since Mistral open-sources its weights. But running inference at scale on your own metal costs more than API calls for most teams. The choice is now: pay the enterprise premium, accept that you're a data contributor, or spin up your own infrastructure.
The Implication
If you're building an agent or application on Mistral's API, audit your tier and your data handling. If you're below enterprise, assume every prompt is teaching the model. That might be fine for generic tasks. It's not fine if you're handling proprietary workflows, customer data, or anything you'd rather not see surface in a future model's training corpus.
The broader pattern is clear. AI companies are converging on a two-tier privacy model: free/cheap users are data sources, enterprise buyers get isolation. Mistral just made that split explicit. Watch for smaller model providers to follow. The era of "we don't train on your data" as a competitive differentiator is closing.