Skip to content

Dashboard and costs

Volume, failure rate, latency and recorded spend for the selected environment and range.

On this page

Both pages answer "is this healthy, and what is it costing" for the environment and time range selected in the header. Every number on them opens the runs behind it.

Dashboard

Available on every plan:

  • Stat row — requests, failure rate, latency and recorded spend for the selected range, each compared with the period immediately before it.
  • Requests chart — trace volume and failures, zero-filled so a gap reads as "nothing happened" rather than "no bucket".
  • Recent traces and recent failures — the newest ten of each, failures first. Both open the exact run.

Clicking a point in the chart drills into the traces for that slice with the time filter applied.

Full dashboard Pro

Below the basics, Pro and Enterprise add:

CardWhat it shows
Cost overviewRecorded spend for the range, with its change
Top models by costThe cost-ranked models behind that spend
Run duration percentilesp50, p95 and p99 for completed runs
Slowest operationsThe slowest models and tools among the most-called groups
AgentsRuns, failure rate and cost per agent
Token usageInput and output tokens over the range

Each card links into the matching Analyze page with the same environment and time filters carried across. Cards with nothing to show are hidden rather than rendered empty; Agents appears only when more than one agent is named, and lists the six busiest plus Other.

A failed breakdown leaves the basic dashboard intact with a retry — one card's failure never takes the page down.

Free plans get one explanation in place of these cards.

Costs

Cost is computed per span from the model, the provider and the reported token usage, including cached and reasoning tokens. On every plan:

  • Recorded USD, average cost per finalised run, input and output tokens and run counts, each with its previous-period change.
  • A zero-filled spend chart across the range.

Pro and Enterprise add breakdowns by model, provider, agent and environment, each with a complete Other total so the parts always add up to the whole, and the 20 most expensive retained finalised traces.

Costs shares the environment and time filters, the sidebar, the command palette and g c. From the dashboard's spend card you land here with the same range; from a trace you return to the Costs view you came from.

What the numbers do not cover

The pages say all of this in place, and it is worth knowing before you act on a figure:

  • Unpriced models. A model without price coverage is counted and timed but contributes no recorded USD. Cost is a floor, not a ceiling.
  • Aggregate rounding. Breakdowns come from hourly aggregates, so totals are rounded.
  • Raw history. The most expensive traces, recent model and tool calls and the Spans explorer read the latest 24 hours within the selected range, because they need span detail rather than aggregates. Aggregate-backed queries cover up to 90 days.
  • Retention. Runs outside your plan's window are gone; their aggregates and usage counters are not.
  • Sampling. Above the monthly allowance, non-failed traces may be sampled. Sampled traces keep their counters and metrics but no span detail.

Budgets, per-tool and per-tag cost groups, pricing management and optimisation suggestions are not built yet.

Latency

The dashboard's P95 card and the command palette both lead into the Latency tab of the Spans explorer: completed-run percentile comparisons and trends from aggregate histograms, plus span-kind, model/provider and tool groups compared by operation duration. The slowest retained runs open on their longest operation, skipping the root run wrapper, and keep your investigation when you come back.

Percentiles are approximate, missing timings are named rather than assumed, and critical-path analysis and regression detection are not built yet. See Models, Tools, Spans and Prompts.