Usage

Where to find it

Sidebar → Usage. Pick a billing period from the selector at the top right.

Usage page with the period selector, headline totals and the top agents chart

Panel	Tells you
Period totals	Tokens and spend for the selected month.
Top agents	Which agents consume the most, ranked.
Trend	How consumption is moving period over period.

Every model step inside a run records four token counts and a cost figure:

Measure	Meaning
Input tokens	What was sent to the model: instructions, history, retrieved passages.
Output tokens	What the model generated.
Cache read tokens	Reused from a cached prompt. Usually much cheaper.
Cache write tokens	Written into the cache for later reuse.
Cost (USD)	Recorded per step, to six decimal places.

Because every member has their own gateway key, spend is attributable to the person who ran the agent, not just to the organization.

Lever	Effect
Use a smaller model where quality allows	The single largest saving, usually by a wide margin.
Shorten instructions and skills	They are re sent on every single turn.
Attach fewer knowledge bases per agent	Each one adds retrieved passages to the input.
Cap output length in the instructions	Output tokens cost more than input tokens on most models.
Set a model allowlist	Stops expensive models being picked by default. See Model Gateway.
Look for retry loops in traces	A tool failing and retrying three times costs three times.

Use Traces for accurate figures today:

Anyone with trace: read can do this across the whole organization without being able to read conversation content.