Training data disclosure

Yes — unless you opt out. Prompts and completions sent through our endpoints may be used to train our own models. Opting out takes effect on your next request and stops all future content retention for your account.

Open dashboard → "Train on my traffic" switch. Usage is disclosed on our plans page and in the signup terms before your first key is issued; you can flip it anytime without any effect on service quality or price.

Our end-state — stated on the home page — is domain-specific small models that answer coding questions well enough, cheaply and privately enough that customers with sensitive data never have to ship their prompts to frontier labs. The scarce input for that is honest, well-routed, real-world coding traffic — which our users generate. We would rather fund that with your voluntary consent than by selling the service short with fabricated research claims.

• prompt and completion text of your API requests (including tool-call payloads);
• routing metadata (which model tier answered), timing, and cache statistics;
• you are not identified in any model output by name, account, or code you submit; we do not publish datasets containing your content.

• your request content stops being stored by us (metadata for billing continues — it is needed to run the service);
• content already used in a model training run cannot be un-learned — deleting the source material from our stores does not retroactively remove gradients from trained weights. This is the honest limit of opt-out, and we state it the same way regulators are starting to require;
• deleting your account erases your retained content within 30 days.

Each cohort report discloses: consent rate in cohort, share of traffic used for training, and upstream retention terms we rely on. No per-user rows are ever published.