Usage and costs
What was spent, by whom, on what, and why the number is an estimate.
Usage in the console. A date range (today through 90 days) and four headline figures: total requests, total tokens, estimated cost, error rate.
Cost figures are estimates, computed from the pricing you configured on each model. For amounts actually billed, look at your provider's dashboard. If your organisation runs on credits none of that applies: what you are charged is what the model actually cost, reported on the Credits page.
Which means the estimate is only as good as the pricing you entered. A model with no price contributes requests and tokens but nothing to the cost, so the total reads low and nothing says so on the headline. That is the failure mode to watch for: an unexpectedly cheap month is more often a missing price than a quiet one.
The breakdowns
Daily usage over the range, for shape rather than precision.
By model shows where the volume is. Toggle between display names and raw ids when you need to match something against a provider's own reporting.
By user is the honest one. It shows who is actually using this, which is what you need at renewal and what tells you whether a rollout landed or fizzled.
By project attributes API traffic. Projects that have been deleted are grouped rather than dropped, so historical totals still add up.
Cost by model is the same data arranged for the question "what is expensive?".
Errors and fallbacks
Worth more attention than it usually gets, because it is where a pattern shows up that no single user can see.
Hard failures by category: rate limits, context too long, authentication, server errors, timeouts, unsupported features, quota exceeded, and "all models failed".
Fallback triggers by category is the interesting half. These are requests that succeeded, on a different model than intended, because the primary failed. The user was told at the time, by a banner naming both models, but one banner rarely prompts anybody to report anything. Here and on the Dashboard alert is where the rate becomes visible.
Read the categories as instructions:
| Category | Usually means |
|---|---|
| Rate limit | Your provider tier is too small for the load |
| Context too long | A context window is set wrong, or people are having very long conversations |
| Authentication | A key has expired or been revoked. Fix now |
| Quota exceeded | Provider spend cap reached |
| All models failed | Your fallback chain is exhausted or absent |
Recent errors lists individual ones with model, message and time, which is what you want when somebody reports something odd at a specific moment.
Per-project detail
Each project carries its own usage tab and optional monthly budget. This page is the organisation-wide view; the project tab is where per-surface questions get answered.