Introduction
Monitoring's cost view and Cost Explorer read the same telemetry and answer different questions. This one asks whether spend is behaving; Cost Explorer asks where it is going. Knowing which you are in saves a lot of circling.
Where to find it
AI Governance > Monitoring > Cost & Tokens.
The tab shows token count (input and output), total cost per day, and cost per trace per day. Volume and errors are on the Traces tab.
Which screen for which question
You want to know | Go to |
Did spend change, and when | Monitoring · Cost & Tokens — total cost per day shows when it moved |
Which agent or model is responsible | Cost Explorer — built for attribution |
What the month will cost | Cost Explorer — it carries the projection |
Whether a spend change is behavioural | Here — cost per trace per day shows whether each run got more expensive. Volume and errors are on the |
That last row is why cost per trace is shown. Total cost on its own cannot distinguish growth from waste. Cost per run can.
Why tokens are shown, not just money
Tokens are what actually happened; cost is tokens multiplied by a rate. Separating them lets you tell apart two situations money alone conflates:
Tokens flat, cost up — the rate changed. A different model, or a shift in the input/output mix, since output tokens usually cost more.
Tokens up, cost up — the agent is genuinely doing more. Whether that is good depends on volume.
Note: Input tokens are usually the larger share, and the most reducible. The biggest single input is almost always context that nobody has audited since it was written. Trimming it is the highest-yield cost change available in most agents, and it does not touch the model choice.
Reading it
Compare total cost with cost per trace — Rising total cost with flat cost per trace is more runs (confirm on the
Tracestab). Rising cost per trace is a change in behaviour, and worth chasing today.Check whether you changed anything — Enabling LLM-based controls or an AI judge adds model calls on every request or session. Both show up here as agent spend.
Then move to attribution — Once you know spend changed, Cost Explorer tells you whose — see Cost Explorer Overview.
"Cost per run is a diagnosis. Total cost on its own is just a number going up."
The standing caveat
Everything here is derived from instrumented traffic, and every gap in instrumentation makes the figure read low. It will not match your provider invoice, and the reasons are all legitimate — see How Cost & Tokens Are Calculated.