Your model bill doubled and someone suggested a cheaper model. It can cost 40% less per token and still cost more per finished task if it fails more often and needs retries. Fullbeam runs both setups on your merged PRs and shows what each one spends per task that passes.
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.
Fullbeam adds up the model and runtime cost of every valid attempt, failures and retries included, and divides by the tasks that passed. Engineer time isn't in that number, and it's labeled that way. A missing cost shows as Unknown, never as zero.
A cheaper model can match your current setup on backend fixes and fail three database migrations. Fullbeam reports those separately, so one average doesn't hide the failures.
A benchmark isn't production. Start the switch on the kinds of work that passed, using your own rollout tools, and check the real bill and results before going wider.
Fullbeam won't show a yearly savings number until your real spend, how much of each kind of work you do, and enough results back it up. Until then the forecast shows Unavailable.
Move the kinds of work where the saving holds. Keep migrations, auth, or anything with too few tasks to judge on your current setup until the results say otherwise.