See and control cost

Shows what agents spend by agent, function, model, provider and user, and brings it down without making them worse.

Ask your AI client#

What did our agents cost over the last 30 days? Break it down by agent and by model, and tell me what our acceptance tests cost.

What you get back#

Spend from the same ledger the Studio reads, so the two never disagree. Every model call is recorded with the tokens the provider reported, priced at that model's list rates. Every breakdown (by agent, function, model, provider, user or interaction) adds up to the same total, and the interaction breakdown separates real use from test runs. Each environment is counted on its own, so a busy test environment is real spend.

What to watch#

Weekly, name the top three agents and note each one's Avg $/call and Avg input tok/call. Read them again a few days after any activation or model change.

What moves	What it usually means	What to change
Avg $/call up after an activation	The new version sends more, or runs on a larger model	Read Avg input tok/call, then compare the versions
Avg input tok/call up	An operation returns more than it used to, or the prompt grew	Return less
Cached share falls	Something changes at the start of every request	Ask whoever built the agent whether it can move
Test runs are a large share	Acceptance tests run often, many times each	Check Optimize → Runs

Bring it down#

In rough order of savings:

Prove a cheaper model still passes. Run an experiment on a smaller model. No regression means switch, and you have the run to show anyone who asks.
Turn fixed work into a function. Comparing invoice lines to PO lines is arithmetic. As function steps it costs less and stops varying.
Use machine checks instead of judged gates where the condition isn't about quality. A judged gate is a model call every time the phase is checked.
Return less. An operation that returns a whole record when four fields are needed sends the rest to the model.
Schedule for the rate the data changes. Hourly against data that updates daily costs 24 times what it needs to.

What you decide#

What each agent is worth, and the model it runs on. Better answers at twice the cost can be the right trade, if you make it on purpose.

To alert on spend, turn on Send costs in the OpenTelemetry export and set the threshold in your own tooling. The figure is priced from list rates, and your provider's invoice is the final word.

Done when#

You can name your three most expensive agents and say why.
You ran one model comparison and either switched or can say why not.
You know what each scheduled agent costs per week.

In the Studio#

Monitor → Cost, the default tab, shows Total spend, Metered calls, Avg cost / call, Avg input tokens / call, Tokens in and Tokens out, with Spend over time and a breakdown table for each axis. The built-in agents count against your org only if Run the built-in agents on our own key is on in Settings → Governance.