Quick answer
Cost per developer is total AI spend divided by active developers for the same period, with fixed seat cost and variable usage reported separately. Segment by workload before drawing conclusions, because an agent-heavy developer and a reviewer have different needs and different costs.
The calculation
cost_per_dev = (seats + variable_usage) / active_developersUse active developers, not headcount. Including people who never signed in makes the number look efficient and hides the wasted seats.
Segment before judging
| Segment | Pattern | Plan fit |
|---|---|---|
| Agent-heavy | Long runs, high tokens | Top tier |
| Feature work | Daily sessions | Mid tier |
| Review and chat | Short interactions | Entry tier |
| Platform | Spiky, project-based | Mid tier plus API |
Use it for sizing
The number should decide seat count and tier mix, not performance. A developer with high token use may be doing the hardest work, or fighting a looping agent; the cost figure alone cannot tell you which.
UsageAtlas runs per machine, so each developer gets their own history and can share aggregates. See team usage.
Two numbers on one slide
Put cost per developer beside a workload measure, never alone. A cost figure without volume invites the wrong conclusion: the highest number might be the person doing the hardest work, or the one whose agent loops. Pair spend with sessions, completed tasks, or tokens per completed task to give the number a direction.
Then split the population into two or three workload groups and report each. Averages across very different jobs hide the only finding that matters: which group needs a bigger allowance and which needs a cheaper default model. The team usage guide covers the grouping.
Questions, answered
How do I calculate AI cost per developer?
Add seat cost and variable usage for the period, then divide by active developers. Report the two cost types separately.
Should high usage be treated as a problem?
Not by itself. Task mix drives usage. Look at cost per completed task and whether the work was necessary.
What is a normal number?
There is no reliable benchmark. Compare your own month over month with the same tools and workload.