Trust & safety

Data retention

By default Utmost keeps zero content: no prompts, no outputs. We keep request metadata, which we need for billing, rate limits and routing.

Content is stored only in three cases, each of which you turn on or ask for, and each with an expiry.

Summary

Data Stored by default? Retention
Request metadata Yes Kept for billing and usage history
Prompts and outputs No Not stored
Audit log (full requests and responses) No; opt in 1 to 365 days, your choice (default 30)
Stored responses (Responses API) No; opt in 30 days, or until you delete them
Batch inputs and results Only while a batch runs Until 7 days after the batch finishes
Safety incident metadata Only on a hard-limit block Kept for trust and safety review
Webhook delivery records Yes, if you use webhooks 30 days

Request metadata

For each request we record:

  • request id, endpoint, key, project and organization;
  • the model you asked for, the model that served it, the route, tier and variant, and any fallback;
  • HTTP status and error code;
  • input, cached, output and reasoning token counts, and whether they were estimated;
  • cost, latency and time to first token;
  • whether it was streamed or part of a batch;
  • your feedback rating, if you sent one.

This is what the dashboard's usage pages, GET /v1/credits and the usage.daily webhook are built from. It contains no prompt or output text.

Opt-in: audit logging

Turn on audit logging in the dashboard's Settings, and choose a retention period from 1 to 365 days (default 30). While it is on, each request and its response are stored with the request id. That includes batch requests and workflow passes. For streamed responses, the stored response is the generated text only, without tool-call deltas. Entries are deleted when they reach the retention period. You can also delete all audit logs from the dashboard at any time.

Turning audit logging off stops new entries. Existing entries stay until they expire or you delete them.

Opt-in: response storage

The Responses API can store responses so you can retrieve them or continue with previous_response_id. Response storage is off by default. With it off, store: true returns storage_disabled. See Responses.

With it on, a stored response holds the conversation up to and including that response. It is deleted after 30 days, or when you call DELETE /v1/responses/{id}.

Batches

A batch has to keep its inputs until they run. See Batches.

  • Each request's input is deleted as soon as that request succeeds or fails.
  • Results, and the inputs of any requests that never ran, are deleted 7 days after the batch finishes.
  • The batch object (status, counts, cost, metadata) stays.

Put nothing in batch metadata that you would not want kept.

Safety incidents

When a request is blocked under a hard limit, we record the organization, key, request id, category, whether the input or the output was flagged, and which classifier decided. We do not record the prompt or the output, even then.

What we never store

  • Prompts and outputs, unless you turned on audit logging or response storage, or submitted a batch.
  • Content seen by the safety classifier. The check runs in memory.
  • Content seen by the routing classifier.
  • Full API keys. We keep a hash and a short prefix.

Abuse reports

Reports sent through /abuse are stored with the reporter's email address, the key prefix and request id they gave, a category, and their description, so trust and safety can act on them.