A cheaper token price does not help if the agent fails more often. Cost per accepted task takes everything you spent on valid attempts, failures and retries included, and divides it by the number of tasks that passed.
Every result links back to the setup, the tasks and the commit it came from. If something wasn't measured, you see that too, instead of a number that looks complete.
Say setup A costs €0.40 per attempt and passes 8 of 10 tasks. That is €4.00 for 8 passes, or €0.50 each. Setup B costs €0.25 per attempt but passes only 4 of 10: €2.50 for 4 passes, about €0.63 each. The cheaper setup costs more per task that passed. A pass means the tests from the merged PR pass. Tokens used or lines changed never count.
Backend fixes might be cheap while database migrations burn through retries. Report cost for each kind of work before you average anything, or one cheap category will hide an expensive one.
The number covers model and runtime cost only. Steering, restarts, CI fixes, review rounds and rework are real costs too. They are reported only when they are measured for the same tasks. Otherwise they are marked unavailable and left out of the number.
If there is too little data on current spend, task volume, what actually ran or which tasks passed, no saving is shown. The estimate stays blank until the data is there, rather than showing a guess.
A new setup is cheaper when it passes the same tasks for less measured model and runtime cost, across repeated runs. That number leaves out engineer time, so it is not a total engineering cost.