LobstackConsole
Performance
How fast your calls are, how often they work, what a finished task costs, and how well your models call tools.
/dashboard/performance reads the same request trace and ledger as Usage, over the same window. Pick a model to narrow the charts to it. The figures are also at GET /api/v1/metrics with a usage:read key.
Latency and overhead
| Figure | What it measures |
|---|---|
| First token | On streamed calls, the time from the Gateway starting the provider's stream to the first token. p50 and p95 per day. |
| Total | From the call arriving at the Gateway to the answer being ready, including the provider. p50 and p95 per day. |
| Lobstack overhead | Total minus the time spent waiting on the provider, every attempt included. The median over successful calls. |
Overhead covers checking the key and limits, routing, the spend hold and pricing. Writing the ledger row after the answer is not in it. Only successful calls count toward latency: a refusal is fast for reasons that are not performance.
Success rate
The share of calls answered without an error, per day, then by model and by provider. Failures are split by the six classes on Errors & retries, so a quota refusal (yours) is not read as a provider outage (theirs). A call refused before a model was chosen counts in the total, not under any model.
Cost per finished run
Median and p90 cost of the runs your client finished, by the day they finished and by client tag, with the most expensive runs listed. A run is every call that carried one run id; see Runs for the headers.
Tool-calling evals
The latest scores from the public tool-calling suite, for the models you called in the window, and Nex if you sent auto. Until the first run is published it says so.