Act on this, mostly signal

Four Claude Code token savers tested, and only one cut cost on large jobs

What to do

Test a token saver on your own large tasks before keeping it, since small tasks can cost more.

Why

On hidden tests, ponytail cut median cost 34 percent on a large feature but added 20 percent on a small bug. Caveman moved cost about 3 percent.

Read the source at reddit.com

What the research found

Reddit test of 4 Claude Code token savers with hidden tests (Sonnet 5.5, Claude Code 2.1.289): ponytail v5.3.2 cut median cost 34% on the large WeasyPrint feature but added 20% on a small bug; caveman v4.0.1 moved cost about 3%; superpowers v7.0.0 added cost; answer-me-with-html never activated. Small sample, API-rate estimates.

Claims and how well the evidence supports them

Ponytail v5.3.2 cut median cost 34% on the 8-file WeasyPrint feature task 98%
Ponytail v5.3.2 added 20% cost on the small humanize bug task 98%
Caveman v4.0.1 cut output 21% on the bug but median cost moved only 3% 87%
answer-me-with-html v0.5.5 was never activated in 6 runs 99%

Evidence

reddit.com https://reddit.com/r/ClaudeAI/comments/1xdqzzb/i_tested_4_claude_code_token_savers_on_real_tasks
github.com https://github.com/DietrichGebert/ponytail
github.com https://github.com/JuliusBrussee/caveman

Actionability rated medium.