Zero data retention

The following are processed in volatile memory for the duration of your request and discarded when the response has been delivered:

• Prompts and messages, including system prompts and conversation history you send.
• Images and any other attachments in a request.
• Tool and function definitions, tool-call arguments and tool results.
• Model outputs: completions, streamed tokens, reasoning text and tool calls.
• Any file, code, secret or personal data contained in the above.

This applies to every request on the lane. There is no default logging mode to switch off, no sampled retention, and no staff review of content. We do not train models on your data and we do not sell or share it.

To run the service and bill you correctly we record metadata about each request. Metadata describes the request; it never includes the request.

Aggregate figures built from this metadata, such as the tokens-per-second and cache-hit numbers in our beta report, contain no content and cannot be traced back to a prompt.

Follow-up turns are fast because the lane keeps the token-level state of your recent context in accelerator memory. This cache is what makes a 200K-token conversation resume in under two seconds.

• Cache entries live only in the working memory of the machine serving your lane. They are never written to persistent storage.
• Entries are evicted automatically as new work arrives, typically within minutes to hours, and are cleared entirely whenever the serving process restarts.
• A cache entry can only be reused by a request that presents the identical token prefix, which in practice means your own follow-up turns. Cached state is not readable through the API and cannot be exported.

• Requests and responses are encrypted with TLS between your client and our endpoint, and between the components that serve your lane.
• Request bodies are not written to disk at any hop. Our gateway inspects only what routing needs: your API key, the model id, and request parameters such as token limits.
• If a request is routed to standby capacity that we operate, the same commitments on this page apply to it.

Zero data retention covers what you send to the model. Information about your account, such as your email address, your beta application answers, your bookings and your payment records, is held as described in the privacy policy. Card details are handled by our payment processor and never reach our systems.

• We enforce our acceptable-use terms using account and metadata signals only (for example, request volume, error rates and payment signals), never by reading content.
• Because content is not retained, we cannot produce it in response to a legal request. We may be required to disclose the account and billing metadata listed above.
• If we ever needed to change this policy, the change would be announced in advance and would never apply retroactively to requests already served.

Send a request and open your dashboard: the usage view shows request counts and token totals for your booked hour and nothing else, because nothing else exists. If you have a compliance questionnaire or need a signed statement of this policy, ask us on Discord and we will provide one.