Usage and Request Logs
Read workspace usage, performance, costs, and request traces from the current Token Usage page.
Open Token Usage (/usage) to inspect the currently selected personal or organization workspace. This differs from Admin Panel → Insights, which is platform-wide and restricted to administrators.
Filters and metrics
The date range and API-key selector are global filters. They update summary totals, charts, performance samples, and request-log data. Endpoint and model selectors narrow the request-log table; they do not recalculate the summary cards above it.
The page reports requests, tokens, spend, TTFT, throughput, and latency where the runtime recorded those fields. Missing performance data is unavailable, not zero. Daily charts and model distributions use aggregated usage records, while performance values and trace details come from request logs.
Dates follow the administrator's system-time mode. In viewer-local mode the browser sends its timezone offset so day boundaries match the viewer. Changing a filter cancels obsolete requests, and Load more retrieves additional server-filtered rows rather than only re-filtering the first page.
Request-log table
Rows include time, API key when present, request ID, endpoint, model, status, TTFT, throughput, latency, tokens, and cost. Column headers support sorting; API key, endpoint, and model headers also expose filters. A deleted key can remain on historical rows with a deleted marker.
Select a row to load the available request trace. A trace can be absent when the request has no request ID, retention has expired, or the trace backend is unavailable. Failed requests may have incomplete token, timing, or cost fields.
Logs API
GET /api/logs is an authenticated API-key view of request logs. It requires start_time and end_time, accepts Unix timestamps or parseable date-time strings, and supports optional model and apiKey filters. apiKey must be a valid user API key; filtering is applied by key identity rather than matching a displayed name.
Use an explicit timezone in date-time strings. The endpoint enforces the exact second-precision window even when storage queries use date partitions, and rejects an end time earlier than the start time.
Interpretation
- Usage aggregation, request logs, and wallet settlement can complete at different times.
- Endpoint/model filters on the table do not change the summary cards.
- Performance charts use valid samples and are not billing ledgers.
- Use a request ID to correlate Studio logs with runtime or upstream-provider logs.