Usage Analytics and Platform Insights
Review platform-wide requests, tokens, spend, performance, user rankings, and request logs from the Admin Panel.
Admin Panel → Insights provides platform-wide observability across users and organizations. Administrators can track usage, performance, and spend trends, then use request logs to investigate anomalies.
This page is available only to accounts with the platform Administrator or Super Administrator role. Regular users cannot open the Admin Panel.
Open Insights
Dates follow the platform system-time setting. Custom ranges are constrained to supported past dates. Changing the range reloads aggregated usage and request logs for that period from the server.
The header also reports when the latest usage rollup finished. If that job failed, an error state is shown. Live requests may still appear in the log while summary totals remain stale.
Summary metrics
| Metric | Meaning |
|---|---|
| New registrations | Users created during the selected period |
| Requests | Total request count in platform usage rollups |
| Total tokens | Combined input and output tokens |
| Total spend | Cost recorded in usage and wallet-backed records |
| Average throughput | Average tokens per second among requests with a measurement |
| TTFT P50 / P99 | 50th and 99th percentiles for time to first token |
| Latency P50 / P99 | 50th and 99th percentiles for full request latency |
When no measurement samples exist, TTFT, throughput, or latency is shown as unavailable rather than as a real zero value.
Trends and distributions
Insights displays the following for the selected period:
- Daily registrations.
- Daily request traffic and token volume.
- Daily spend.
- Most-used models ranked by request count.
- Spend distribution by model.
- A TTFT-versus-throughput performance scatter plot.
The popular-model and spend charts show the leading models with recorded usage, not every model published in Model Zoo. The performance scatter plot uses at most 250 recent valid samples from the available data. Use it to inspect distribution and outliers, not as a complete list of requests in the period.
User usage ranking
The ranking aggregates the currently loaded request logs within the date range by user:
- Request count
- Total tokens
- Total cost
- Average latency
Select a column heading to change the sort order. The view shows up to the top 50 users. Use it to identify high-usage or high-latency accounts, not as a replacement for finalized billing records.
Platform request logs
The request log table includes:
- Time, user or organization, and API key
- Request ID, endpoint, model, and status
- TTFT, throughput, total latency, and available latency stages
- Token count and cost
Filter by endpoint, model, user or organization, and API key. Page sizes of 10, 20, 50, and 100 are available, and older records can be loaded on demand. When the live connection is available, new requests are added automatically.
Select a row to open its details. Studio uses the Request ID to load available trace data. Details can be incomplete when the request has no Request ID, its trace is past the retention period, or trace loading fails.
Interpret the data
Usage rollups, live request logs, and wallet charges can settle at different times. A new request may first appear in the live log while its cost is still settling. Wait for rollup refresh before final reconciliation.
- Request, token, and spend totals use period rollups; performance percentiles and user rankings require request-log samples.
- Requests without TTFT or throughput measurements are excluded from those calculations.
- Historical rows can retain a “Key Deleted” marker after an API key is removed.
- Failed requests may not contain complete token, performance, or cost fields.
- Insights is an operational platform view. Investigate RAG sources and conversation behavior from the relevant custom model's request logs.
Recommended investigation flow
- Check requests, tokens, and spend for unusual spikes.
- Compare TTFT, latency, and throughput to identify queuing, generation, or end-to-end issues.
- Narrow the scope with model popularity, spend distribution, and user ranking.
- Filter request logs by endpoint, model, user, or API key.
- Open a specific Request ID and inspect its status, timing, and available trace.
- When necessary, continue with API-key, upstream-model, wallet, or system-model configuration checks.