Claude Code Cuts Token Use 45% by Switching 5 Tasks From High to Medium

1 articles · Updated · XDA Developers · Aug 8

Across five coding tasks, Claude Code used about 14,300 output tokens on Medium versus roughly 26,000 on High, a 45% reduction after a single effort-setting change.
The test reran the same prompts on the same untouched codebase, isolating effort level as the only variable while covering bug fixes, search, refactoring, test coverage and performance debugging.
Task-level savings ranged from 25% on a simple bug fix to nearly 48% on building a search feature, with other runs showing 54%, 40% and 39% lower output-token use.
Medium still completed all five jobs successfully; in two cases both settings found the same under-tested function or performance issue, applied the same fix, and finished with all 12 tests passing.
The result suggests Medium is a better default for routine work, while High, Xhigh and Max remain better suited to harder debugging or architectural tasks where extra reasoning may pay off.