The AI Taxonomy.
Terms  /  Measurement  /  Quality-Per-Token
09 · Measurement

Quality-Per-Token

A measure of output quality relative to tokens consumed. Distinct from Token ROI (value) — this measures correctness, completeness, and relevance per unit of cost.

Quality-Per-Token fills a gap that Token ROI leaves open. Token ROI measures value — did the output produce a business result? Quality-Per-Token measures craftsmanship — was the output correct, complete, and well-executed? You can have high Token ROI (the feature shipped) with low Quality-Per-Token (the code has technical debt, missing tests, and poor error handling). Quality-Per-Token catches the difference between "it works" and "it works well."

Example
Two agents complete the same task. Agent A: 3,000 tokens, produces a function with correct logic, proper error handling, comprehensive tests, and clear documentation. Agent B: 8,000 tokens, produces a function with correct logic but no error handling, minimal tests, and no documentation. Both functions "work" — Token ROI is similar. But Agent A's Quality-Per-Token is dramatically higher: better output at lower token cost. Measuring Quality-Per-Token reveals that Agent A's system architecture (better context, better prompting) produces superior craftsmanship per unit of spending.

More in Measurement

Context Efficiency

Useful retained state per token consumed.

Token ROI

The measurable value produced per token consumed. The business metric for Context Efficiency.

Context Utilization Rate

The percentage of input context that the model actually uses for its output. Low rate = waste.

Recall Cost Ratio

The cost of retrieving stored context vs. the cost of re-generating it from scratch.

Waste Ratio

The proportion of tokens consumed that produce no useful output. The metric behind Token Bleed and Token Burn.

Iteration Velocity

The speed at which a human-AI pair can move from idea to tested output.

Recovery Cost

The time and tokens required to get a session back on track after a failure — State Loss, Execution Hallucination, or Plan Drift.