Why You Should Measure Tokens Per Successful Task, Not Per Request

Most developers track LLM token usage per API request, but the more accurate cost metric is tokens consumed per successful task completion. Every failed attempt that triggers a retry still bills tokens, meaning three attempts for one answer triples the actual cost of that deliverable. The problem compounds significantly with AI agent loops, where a step retrying three times inside an eight-iteration loop can generate up to 24 times the expected token spend. A developer writing for DEV Community argues this blind spot causes free-tier LLM usage to appear cheap initially before quietly becoming a major pipeline expense. To address this, the author shares a lightweight Python utility that tracks token attribution across all attempts per task, exposing two key ratios: total tokens per successful task and retry amplification rate.