Twelve more model providers, and where your prompt goes

The platform upgrade added twelve inference providers you can select: Mistral AI (FR), Groq, Cerebras, SambaNova, NVIDIA, Fireworks, Together and Ofox (US), Requesty (NL), SiliconFlow (CN), and Replicate and fal.ai (US, for image and video). OpenAI and OpenRouter now cover both direct selection and the routing pool.

More providers means more places a prompt can land, so each is named individually in the privacy policy and in the machine-readable processing register, with its country and its legal basis.

Two things worth knowing before you pick one. The model you choose is not always the provider that answers: on Auto and on the router, a request may be served by a different provider than the one selected, and failover can move it mid-request. The provider that actually served the turn is named on the answer. And a text-only model gets images described by a second provider, which is a separate company seeing your image, so we badge it where it happens.

The register also gained the memory engine, which is zero-knowledge and keeps keys in your browser, with the caveat stated plainly: anything you paste into a prompt is in the prompt. Lose both the passphrase and the recovery code and it is gone. The browser extension and the AI bridge are off by default and consent-gated. Abuse prevention processes IP, user agent and request metadata for 30 days, with aggregates for twelve months.

Three items went to counsel rather than being settled: the China transfer basis, the wording around browser-AI terms risk, and cross-border safeguards for the new US providers. Full list in the privacy policy.