To answer a prompt, we send it to the models you chose. We also send the conversation so far and any files you attached.
Each model's host receives this data under its own jurisdiction:
- Infomaniak, Switzerland: open-weight models;
- Mistral AI, France;
- OpenAI, Anthropic, Google and xAI, United States;
- Stability AI, United Kingdom: image generation.
The Models page shows the host of every model. A prompt leaves Switzerland only when you send it to a model hosted elsewhere.
Switzerland recognises the United Kingdom and France as providing adequate data protection. It recognises the United States only for companies certified under the Swiss-U.S. Data Privacy Framework. Where a provider is not certified, we rely on the standard contractual clauses in its API terms.
Each provider's own API terms decide what that provider keeps.
Voice input goes to the host of the transcription model in the same way.
Our server logs store the model, the token counts, the timing and whether a request succeeded. They never store what you wrote or what a model answered.