Cost per verified outcome.
The cost of the calls plus the execution bill, divided by the outcomes your own eval graded as passing. Not the outcomes someone declared.
What it is
Add up what a set of tasks cost: the model calls, and the bill for running them (compute, storage, network). Divide by the number of those tasks that an eval graded as a pass. The eval is yours, written against your definition of done. A task with no grade does not count in the denominator.
Declared is not verified
Many cost tools take the outcome from the caller: the code that ran the task reports SUCCESS or FAILED. A caller that asserts SUCCESS counts a failed task as a pass, and the cost per outcome falls without the work improving.
Example, not a measurement
Cheaper per call, dearer per outcome
A cheaper route can fail more often. Every failure is paid for and returns nothing. Fail enough and each pass costs more.
Example, not a measurement
Compute it yourself
(call cost + execution cost) / tasks your eval graded as pass
Use the same task set for every route you compare, and count only tasks the eval actually graded. If you want to try your own numbers, the calculator does the arithmetic and says which field it still needs.
What Margin does with it
Margin measures this number from your calls and your eval results. It promotes a cheaper route only when measured parity holds and the cost is lower. When parity slips, it reverts. No customer has supplied a value for an outcome yet. So we report cost per outcome, not value per outcome, and say so wherever the number appears.