Skip to main content
Open Activity in the console to inspect inference requests for the project selected in the sidebar. Choose a 7-, 30-, or 90-day period, filter by model or source, and open a request for more detail. Switch projects to view another active project’s requests; billing credits remain shared by the workspace.

​
What is recorded

Each admitted inference request receives a durable record before it is sent to the backend:
  • Request ID, model, endpoint, and API-key reference when applicable.
  • Source: API or Playground.
  • Status, latency, HTTP outcome, and a safe error code when available.
  • Provider-reported input and output token counts.
  • Settled cost when billing can determine the charge.
The usage record does not store prompts, generated responses, or full API keys. The inference backend still receives your request to generate its response; this is not a claim about the upstream provider’s retention policy.

​
Request statuses

If final recording is unavailable, the original pending row remains for investigation. Do not interpret an old pending row as a successful or free request. Authentication failures, invalid request bodies, and requests rejected before admission do not generate inference usage rows. Model catalog reads, credit reads, and token-counting calls do not generate usage rows either.

​
Token counts

Tokens are recorded from provider usage fields. Missing usage is displayed as unknown, not as zero. An explicit provider-reported zero is a valid count. For a customer cancellation observed by the gateway, final reported usage is charged normally. If final usage is unavailable, the reservation is released with no customer charge and token counts remain unknown; Limitless absorbs the unmeasured supplier cost. A released reservation is distinct from provider-reported zero usage. Timeouts, provider failures and platform shutdown can still require reconciliation. For chat, input and output correspond to prompt_tokens and completion_tokens. For Responses, they correspond to input_tokens and output_tokens. For Messages, recorded total input includes ordinary input, cache-write input, and cache-read input; cache reads are also recorded separately for the discount. The provider’s output total remains the output count. Reasoning tokens may be included in provider output totals; displayed text length is not a reliable token or charge estimate.

​
Summaries and exports

Summary totals cover all matching records in the selected period, including records beyond the current table page. Daily buckets use UTC. The CSV export downloads the current page of records; it is not an export of the entire history. Use the response’s X-Request-ID to find the matching row when investigating a request. When contacting your beta operator, provide that ID, approximate time, and model without sending your key or private prompt. See billing for how reservations and unknown usage affect available credits.

​
Organization reports

The Administration API provides paginated token and cost reports across your organization’s projects. Use UTC Unix timestamps and retain filters while following next_page. Usage reports aggregate recorded token counts; cost reports aggregate known finalized costs. Check reserved credits separately for requests whose final usage is still unknown.