The efficiency of language model usage, measuring how many tokens (words/subwords) are consumed to achieve a task.