Running the numbers
You bring your own Claude credential and the team spends it. That is the whole billing model — nothing here provisions a subscription for you and nothing meters you. So the only honest question is: what does a month of this actually spend?
Below is that figure from our own run, and a box to scale it to yours. It is a real measurement rather than a price list, which means it comes with a caveat big enough that it is printed next to every number rather than in a footnote.
For scale: this project merged 182 pull requests into its public repo in its first week open.
Why the second number is so much bigger
Because cost per merge is not normally distributed and pretending otherwise is how estimates get embarrassing. Most changes are small and cheap. A few go four rounds with a reviewer that keeps finding real problems — and that is the system working, not failing, but you pay for every round. The spread between those two columns is the honest range, and the truth for a given month is somewhere in it.
What these numbers do not cover
They are measured over — of merges, not all of them. Spend is only counted where it resolved to specific agent runs — — merges of —, between — and —. The rest carry a zero placeholder that would drag a naive median to nothing, so they are excluded rather than counted as free. If the uncounted merges were systematically cheaper or dearer than the counted ones, every figure here moves.
They are also this codebase, these models, this team configuration. A repo with slower tests, longer files or a stricter reviewer costs more per merge; a quieter one costs less. Treat the median as an order of magnitude that has actually been observed, not as a quote.
The method behind the exclusions, and every other series the record publishes, is on the record.