Mistral Vibe uses user data for training by default, while Vibe Enterprise is opted out

Our organization chose Mistral partly for European privacy controls, but Team users appeared opted into prompt training by default and we could not centrally disable it for a time. Mistral removed some testing prompts after we complained, and its documentation conflicted with what users saw for several days.

I think the terminology is backwards here. "Opt in by default" normally means training is disabled until I enable it. Do you mean users were opted in, with an opt-out required?

I can disable training for both Vibe and Console/API across my whole organization in the admin console, and I am not on Enterprise. I turned it off when I created the account and it has stayed off.

If an organization cannot centrally enforce an opt-out, a per-user privacy checkbox is not a meaningful Team control. Employees handling sensitive IP or GDPR-regulated data should not have to remember an individual setting.

That concern does not match what I see now. The admin console has organization-wide toggles for Vibe and for Console/API, so a Team admin can disable training for both.

I do not see why I would use Mistral when other models outperform it and many can run locally. The European angle is one reason for some companies, but it does not settle the privacy issue.

I use Mistral OCR and its smaller models for simple translation, and they work well for those jobs. I also value a European provider that can answer regulatory questions about training data, even though downloadable weights are not the same as open source.

Calling it spyware ignores the opt-out setting. I have disabled use of my data for training, though any hosted service still processes prompts on somebody else's computer.

My concern is that Mistral was marketed as a European alternative that would avoid this sort of practice. Training on user feedback may help model tuning, but the default and the organization-level enforcement matter.

I was about to pay Mistral after leaving Claude, but I do not want data collection in a paid product. I may use Kagi Ultimate for the inference interface, hoping it adds some anonymization before providers collect data.

This submission is overstating the change. Mistral has long had an admin toggle that permanently disables training, prior opt-outs remain honored, and organizations can set it for their users. Kagi Ultimate still routes much of its traffic to closed OpenAI and Anthropic models.

I use Mistral's GLM-5.2 cheaply without training on planning prompts because I remembered to disable the Studio privacy setting. My basic Vibe subscription still shares data, so I keep sensitive high-level planning in Studio rather than Vibe.

If data collection is the actual problem, switching among Claude, Mistral, and other hosted providers does not solve it. Run models locally.