Apps
AI coding cost per project is useful and fragile
Useful because it names the expensive repository. Fragile because the attribution was your choice.
AI coding cost per project is a useful number and a fragile one. It is useful because it tells you which repository consumes the budget, and fragile because it depends on a decision you make rather than a fact the tool records. Both halves deserve attention before anyone quotes a figure in a planning meeting. The UsageAtlas guide walks through the mechanics.
Start with the decision
Which project owns a session that touched three repositories? There is no correct answer, only a rule.
Three common rules, in order of defensibility.
The session root. The repository where the session started owns the cost. Simple, reproducible, occasionally unfair.
A written split. Shared work is divided by an agreed ratio, such as lines changed or time spent. More accurate and harder to maintain.
Platform. Work that serves every project is charged to none. Honest for tooling, easy to abuse.
Pick one, write it down, and apply it every week. A rule that changes with the result is not a rule.
Then the buckets
| Bucket | What lands there |
|---|---|
| Direct | Feature work in one repository |
| Shared | Cross-repo refactors |
| Platform | Setup, config, tooling |
| Separate | Abandoned sessions |
| Unallocated | Sessions with no working directory |
Watch the last one. If it grows, the per-project figure is drifting toward fiction, because cost is being attributed to projects that did not incur it.
The arithmetic
Multiply tokens by model rate for each model in the session, sum, and divide by completed tasks in that project.
The rate matters more than people expect. A project that ran on a heavier model for routine edits will show a high cost, and the fix is routing rather than a bigger budget.
Why the fee stays out
A subscription is a fixed monthly cost. Charging it to projects makes each project carry an equal share of something it did not cause, which destroys comparability.
Report two lines: tool cost, and usage cost. The second is the one that moves when behaviour changes.
Weekly, not monthly
A monthly average hides a refactor. A weekly figure shows it, and shows it while the person who ran it still remembers why.
That timing is the whole benefit. A number that arrives a month late cannot change the next sprint.
Using it well
Three uses survive scrutiny.
Routing. Point the expensive repository at a smaller model for routine edits.
Estimation. Base the next migration on the last comparable one.
Comparison. Measure the same project in two tools over two weeks, with the same task definition.
FAQ
Is a proportional allocation of shared work honest?
Yes, when the ratio is stated and stable. Silent allocation is the dishonest version.
Should unallocated work be distributed?
No. Distributing it makes every project slightly wrong and hides the gap in the data.
Can the tool do the allocation for me?
It can group by project when a working directory is recorded. The allocation policy is still yours.
The short version
Choose a rule, keep the unallocated bucket visible, and report usage separately from fees. The rest is multiplication.