Back to AIBriefs
AnalysisDevelopers

Analysis finds rtk, headroom, caveman token savings far below claims on real workloads

Analysis finds rtk, headroom, caveman token savings far below claims on real workloadscodepointer.substack.com

Tests on a 614M-token corpus of Claude Code sessions show the three tools together saved only 3.7% of spend, versus claimed 60-90% reductions. Each tool targets a different stream—headroom compresses cache tail, rtk shortens outputs, caveman prunes tool calls—but real-world results are modest.

··Discuss
Jun 18, 4:16 PM
Analysis finds rtk, headroom, caveman token savings far below claims on real workloads — AIBriefs