Cost per Completed Task

Simple Definition

Cost per completed task is the total cost required for an AI system to produce one accepted task result, including model usage, retries, tool calls, agent passes, and verification.

This is a practical metric used in real agent workflows rather than an academic term, but it describes a genuinely useful way to think about AI spend.

Plain-English Explanation

A model’s price per million tokens tells you almost nothing about what a task actually costs to get right. A model charging less per token can still be the more expensive choice in practice if it needs more attempts, generates more retries, makes more tool calls, or requires heavier review before its output is usable. Cost per completed task adds all of that up: inference cost, tool and API cost, the cost of retries, verification cost, and human review cost, then divides by the number of results that actually got accepted.

Example

Model A costs $0.10 per attempt and succeeds 50% of the time. Model B costs $0.16 per attempt and succeeds 95% of the time. Priced per attempt, A looks cheaper. Priced per completed task, accounting for how many attempts it typically takes to get one accepted result, B is often the better deal, since A needs roughly twice as many attempts to land the same number of successes.

How It’s Different From Task Budget and Output Token Efficiency

These three are closely related, and it’s worth keeping them straight. Task Budget is a forecast made before work starts, an estimate of the tokens, time, and effort a task is likely to need. Output Token Efficiency measures useful work per output token, a ratio focused specifically on token spend. Cost per completed task is the realized, after-the-fact number: total dollar cost, across every input that went into getting one accepted result, not limited to tokens and not an estimate. A task can have an efficient token ratio and still have a high cost per completed task if human review time or tool calls dominate the total.

Why It Matters

Judging models purely by sticker price leads teams to pick the option that looks cheapest and quietly costs more once retries, review, and failed attempts are counted. Cost per completed task forces the comparison to happen on the number that actually matters: what it costs to get one usable, accepted result out the other end.

Task Budget, the pre-task estimate this metric can be measured against after the fact
Output Token Efficiency, a related, narrower ratio focused specifically on tokens
Model Routing, a decision this metric can help inform
Model Escalation, a mechanism that trades a higher per-attempt cost for a lower cost per completed task on hard cases