Why it matters
Researcher Denis Shiryaev pinned Claude Code 2.1.201 and Sonnet 5, then measured the only tokens rtk actually compresses: uncached input barely moved (3.2%, p=0.23). More turns (+13.8%) and cache reads (+14.3%) turned a ‘saver’ into a small, consistent tax.
Tokenmaxxing read
Lesson for token-cutters: rtk rewrites ~a third of Bash calls, but Read/Grep bypass it entirely, capping savings near 3% of input. Replaying 83 transcripts predicted that ceiling for ~$0 before any paid run. Measure against real bills, not a counterfactual that never existed.
Source takeaway
Quality was a wash—71 of 80 low-effort tasks tied, sign test p=1.0—so rtk didn’t break code, it just billed more. It’s Shiryaev’s second debunk after the ‘caveman’ skill (9% saved, not 65%). Verify token-saver claims against instrumented baselines.
