A frequently overlooked fact is reshaping the optimization thinking for API-style coding agents: cache creation and cache reads account for up to 87% of the Claude Code bill (by a four-component reconstruction view) or 80% (by actual billing). Sarel Weinberger and Amir Hozez, in arXiv:2607.12161, use 2,908 hash-frozen real Claude Code experiments (2,848 analyzed, 103 tasks, 7 repos, 3 models) to directly refute the intuition that "compressing tool output = saving money". In one arm, after cutting 38% of tool-output tokens, the paired cost actually rose 6.8% (95% CI: +2.8% to +11.3%), the task-level correlation coefficient was only 0.15, and the confidence interval crossed zero. More dangerously, compression destroys the action-evidence chain: on SWE-bench-derived Go tasks, post-compression patch success rate dropped from 27/40 to 15/40, because literal edit anchors got mangled. The paper accordingly proposes success-adjusted billed cost as a new baseline, urging the industry to shift the optimization focus from "compression ratio" to "cache hit structure and evidence fidelity".