Divide attributed spend by landed features, not total spend by the calendar. Teams that trace model costs through coding sessions to the features that actually shipped typically discover unit costs in the hundreds of dollars per feature, along with a long tail of spend that produced nothing and is nobody’s fault until it is visible.
The problem with the monthly total
Most engineering leaders can quote their monthly AI bill. Very few can say what it bought. A five-figure model spend that produced a quarter's worth of shipped product is excellent. The same spend producing three features and a pile of abandoned branches is a problem. The total cannot tell you which one you have.
Building cost per feature
The attribution chain has three links. Sessions carry token usage and model pricing, so each session has a cost. Sessions touch files and produce commits, so they connect to pull requests. Pull requests land in features. Follow the chain and spend stops being a monthly blob and becomes a per-feature figure you can watch over time.
Three practical notes from doing this on real data:
- Count subscription seats properly. Pro and Max plans do not appear in billing APIs, but their sessions still carry token counts. Pricing those at public per-model rates puts every seat on one ledger, clearly labelled as an equivalent-value estimate rather than an invoice.
- Expect a long tail. A meaningful slice of spend attaches to work that never landed. Some of that is healthy exploration. Some of it is a rework loop that needed attention weeks ago. The split only becomes managed once it is visible.
- Watch the trend, not the level. The first month's cost per feature is a baseline, not a verdict. Falling unit cost with steady quality is the signature of a team learning to use its tools.
What this changes in practice
Once cost follows the work, ordinary management questions get answers. Which project is burning premium-model tokens on work a cheaper model handles? Which engineer's two-thousand-dollar month of agents replaced what would have been a fortnight of work, and which one's produced nothing mergeable? Where does a budget flag need to fire before the invoice becomes a surprise? None of these are answerable from a monthly total. All of them are answerable from attributed spend.
Frequently asked questions
What is a good cost per shipped feature?
There is no universal number; complexity varies too much. What matters is establishing your own baseline and trend. Teams commonly land in the hundreds of dollars per feature once attribution is in place.
How accurate is cost estimation for subscription seats?
It is an equivalent-value estimate: session token counts priced at public per-model rates. It is not an invoice, and honest tooling labels it as such, but it is consistent and comparable across the team.
Does cost attribution require developers to tag their work?
No. The chain from session to commit to pull request to feature can be built from data that already exists. Declared priorities improve the importance weighting, but attribution itself is automatic.