Cost-per-Unit-of-Work

Per-token spend is a denominator problem: ten cheap tokens that fail are worse than two expensive ones that succeed. AEOE divides spend by the unit of work it bought — one code review, one summary, one refactor — and scores models on the ratio.