Cost per Successful Task — AI Glossary

Overview

Per-token and per-request pricing hide retries, failed attempts, human rework, and answers that were produced but rejected. Cost per successful task puts the denominator where the value is: define what "successful" means (validated schema, grounded answer, accepted review), then divide total cost — including retries, reranking, and review time — by that count. It matters because a cheaper model that fails more often, or needs three attempts to produce valid output, can be more expensive per completed job than a pricier model that succeeds first time. Common mistakes: comparing candidates on list price alone, excluding retry and review cost, and changing the success definition between candidates.