Dashboard
Model Log
Every model call your app made, newest first — tokens, cost, latency and success or failure per request.
Model Log is the most granular view of your traffic: one row per provider call, newest first, each linking to the run it belongs to.
What it's for
When a number elsewhere looks wrong, this is where you check it against individual calls. It covers every modality — chat, image, audio, embedding and the rest — not just chat.
The columns
| Column | Shows |
|---|---|
| Time | When the call started. |
| Status | Green OK, or a red error kind — hover for the message and HTTP status. |
| Kind | The type of call: chat, embedding, image, and so on. |
| Model | The model, with its provider underneath. |
| User | The user_id your app supplied. |
| In / Out | Input and output tokens. |
| Cost | USD at provider list prices. |
| Duration | Wall-clock time, as your user experienced it. |
| Trace | The first characters of the run id, linking to the full trace. |
Filters
- Status — all, success, or failed.
- Rows per page — 50, 100, 200 or 500, with page navigation and a total count.
- The shared time range at the top of the dashboard.
What to look for
- Every row's User column is empty — your app isn't threading
user_id, so per-user rules can't pool anything. See anonymous users. - A cluster of failed rows at one timestamp — a provider incident or a bad deploy. Compare against the error panels on Models.
- Cost of
$0.00on a real call — the model has no price on file, or the provider returned no usage. Both are itemised on the Unpriced calls tab. - One trace id repeating down the page — a single run making a lot of calls. Open it and check whether a loop rule should be catching it.
- Durations far above your provider's normal — the same calls are charted on the Models tab as p50 and p95.