Research summary and industry perspective

Token reduction is not cost reduction

Four Claude Code configurations, 103 tasks, 2,848 analyzed runs, measured against billed cost. The tested compression configurations changed billed cost differently. This summary distinguishes pooled results, holdout uncertainty, and delivered tool-output tokens. Based on the version 5 paper, with 2,848 runs in the primary analysis.

Based on: Weinberger & Hozez, Token Reduction Is Not Cost Reduction, 2026 / arxiv.org/abs/2607.12161v5 / The full paper is a free download, no registration. This is the research summary and industry perspective: the short version of the study, with the scope and limitations made explicit.