Cost per completed task divides total workflow spending by the number of jobs that were successfully finished and verified. It can include model usage, tools, failed attempts, retries, infrastructure, supervision, review, repair, and cleanup.
The metric connects economics to reliability and quality. A model with attractive cache discounts can still cost more per result if it generates substantially more output, takes extra steps, or needs more retries and human correction.
ELI5
Cost per completed task asks how much was spent before the job was actually finished and accepted. It counts failed attempts and rework instead of looking only at the price of the first AI response.
For example, a cheap model that finishes a tested edit in one try may cost less overall. If it fails three times and a stronger model must repair the work, its low price per request did not produce a low price per completed task.
