Usage

Token counts are tracked everywhere a model is involved:

• Per step — each step in a trace that used a model shows its tokens, split into input and output.
• Per execution — the execution's header totals the tokens across all its steps, including any child executions.
• Per agent or workflow — the rollup table, and each agent's or workflow's own page, total tokens across its executions over the date range.

This lets you find which work is token-heavy, from the whole table all the way down to the individual step.

Duration is tracked per step (tool calls included), per execution, and per agent or workflow as an average. p95 — the 95th-percentile duration, a good measure of worst-case slowness — appears as a tile and a chart on the Activity tab, and as a chart on each agent's or workflow's own page.

Counts and success rates roll up across executions:

• executions over the page's date range (all time unless you narrow it), across the organization or filtered to a project,
• executions over the last 24 hours, per deployment,
• and success rate overall, per agent or workflow, and per deployment.

Together these tell you how busy each piece of work is and how reliably it's completing.

Token usage is priced into a dollar figure using each model's rates, and cost rolls up the same way tokens do — per step, per execution, per agent or workflow, and overall. The Spend tab shows spend over the range compared with the period before, the average cost per execution, spend over time, and your top-spending agents and workflows.

For a model without published pricing (a local model, for example), set per-token rates on its provider connection to enable cost tracking for everything that runs through it.

To put a budget on spend — and block new work when it's hit — see Spend limits.