Telemetry
Investigate usage by cost, tokens, requests, errors, and duration.
Open Advanced → Telemetry to investigate usage behind a total or a change you noticed elsewhere.
Choose a view
Section titled “Choose a view”Select a date range (the last 30 days by default), then the relevant tab. Individual charts offer time granularity controls.
| Tab | What to investigate |
|---|---|
| Cost | Estimated cost over time and by model; read the partial-sum marks on models |
| Tokens | Input, output, cache read, and cache write usage by period and model |
| Requests | Request and session counts, when requests occurred, and the activity heatmap |
| Errors | Recent errors, status codes, error kinds, and retry patterns |
| Duration | Session lengths, and the mean and longest durations |
One turn can contain more than one model request. Session counts, prompt turns, and API requests are different units. Cache reads and writes are also distinct from fresh input and output. A Sessions count is the sessions that sent a request in the period; when the period includes time for which sessions were not counted, it is shown as not counted, not as zero.
The CLI does not record the model of an error, so errors cannot be split by model. The latest errors and the retry pattern cover only the latest days of the selected range, and the retry counts read the newest 500 errors. The screen states the days it covers.
Investigate a spike
Section titled “Investigate a spike”- Match the period used on the original page.
- Select the tab and narrow the date range around the change.
- Check whether one model or a group of requests accounts for the change.
- Review related work in Sessions and any signals in Throttle.
Cost is an API-equivalent estimate, not your invoice. A model marked as a partial sum counts only the requests that have a price. Missing prices, incomplete usage, and partial reads can affect totals or hide a comparison. Read Periods and data states; an unavailable value is not zero.