Definition
Token efficiency is the relationship between quality-constrained outcomes and the model tokens consumed to produce them. It can be reported as accepted outcomes per token or, more readably, tokens per accepted outcome.
The numerator must define the task and acceptance evidence. The denominator must say whether it includes inputs, outputs, reasoning, cache activity, retries, failed runs, and subagents. Use the same task mix, model conditions, and accounting boundary when comparing systems. Otherwise the ratio rewards easier work or hidden consumption.
Efficiency is conditional
An optimization is token-efficient only at a stated quality and risk threshold. A shorter prompt that causes more retries is not efficient. A compressed context that omits a security constraint is worse even if the first answer costs less. Extra inference can improve efficiency when it prevents failure, supports independent verification, or replaces a larger number of bad attempts.
Token efficiency also differs from economic efficiency. Tokens are one input. Tools, infrastructure, latency, review, correction, and incidents can dominate total cost. Cost per accepted durable outcome is the broader measure when those costs matter.
Distinguish it from nearby terms
Prompt caching can reduce billed cost without reducing the tokens a model processes. Token minning is a practice aimed at reducing consumption while holding outcomes constant. Token efficiency is the measured result. Useful intelligence per dollar asks a wider business-value question.
Check your understanding
Version A uses half the tokens but its accepted-task rate falls from 90 to 60 percent. Is it more efficient? Calculate total tokens per accepted result and inspect durability and review cost. Do not infer efficiency from the per-attempt number.