Introduction
Three numbers sit above the chart, and they are not three views of the same thing. One is a rate, one is a ratio, and one is arithmetic performed on the first. Each answers a question the other two cannot, and the ratio is the one that actually diagnoses.
Where to find it
AI Governance > Cost Explorer, above the spend chart.
What each one is for
Figure | Answers | Moves when |
| What is this costing per day? | Volume or unit cost changes |
| What does one run cost? | Only unit cost changes |
| What does a month at this rate look like? | The daily rate changes |
Spend per trace is the diagnostic one
Daily spend rises when you succeed, which makes it a poor health signal on its own. Spend per trace holds volume constant, so a change in it is always a change in behaviour.
Daily rate | Per trace | Reading |
Up | Flat | More of the same work. Usually good news |
Up | Up | Growth and a behaviour change. Two things to investigate |
Flat | Up | Volume fell while runs got costlier. Investigate today |
Down | Up | Fewer, heavier runs — classically a failing path retrying |
When spend per trace rises, the causes are few and all visible in an individual trace: a longer context, more tool calls, retries, or a switch to a more expensive model. See Reading a Trace.
Note: Watch this figure after any change you make to governance or evaluation. Enabling an LLM-based control adds a model call to every request; an AI judge adds one per scored session. Both raise spend per trace without the agent doing anything differently, and both are easy to forget you switched on.
Read the projection as arithmetic
The 30-day projection extrapolates the current rate. It knows nothing about your roadmap, your seasonality, the pilot ending next week, or the launch scheduled for the one after.
It is genuinely useful for two things:
Making a rate legible to a budget holder — A per-day figure means little in a finance conversation. A month does.
Noticing that the rate changed — A projection that jumps says the recent rate moved — worth understanding whether or not the projection proves right.
What it is not is a forecast, and quoting it as one is the fastest way to lose credibility the first month it misses.
"The projection is today's rate with a multiplier. Its job is to make today's rate impossible to ignore, not to predict next month."
Establishing what normal is
All three figures need a baseline to mean anything, and the baseline has to be per agent — a research agent and a classification agent have legitimately different costs per run, and a fleet average describes neither.
The workable habit: record spend per trace for each significant agent when it stabilises after launch, and treat a sustained departure from that figure as the trigger. Compare like with like — if the agent handles several kinds of work, the mix will move the average without anything being wrong. See Run Types.
The standing caveat
All three are derived from instrumented traffic and applied rates, so all three read low against your provider invoice whenever coverage is incomplete, and high wherever your commercial terms beat list price. See How Cost & Tokens Are Calculated.