← Lab
Token savers, tested
Claude Opus 5.5 and Sonnet 5.5 · 2026-11-03
Four Claude Code add-ons promise a smaller token bill: caveman, ponytail, rtk and context-mode. Each one ran the same jobs as plain Claude Code, on Opus 5.5 and Sonnet 5.5 at high effort, and on every job with a right answer, none of them changed the result. The Lab page puts ponytail's builds next to the plain ones: a tower defense game, squishy gummy bears and a landing page.
Some things worth knowing:
- ponytail cost 13% less than plain over five runs of a coding job, and up to 55% less on the builds. Its games were the smallest with both models.
- A two-sentence CLAUDE.md asking for short replies and the least code saved 10%, almost as much as ponytail, with no add-on at all.
- caveman talks less but cost more than plain on four of five Opus jobs.
- context-mode cost 27% more than plain over five runs of the coding job.
Launch
Runs on goosefeathers-lab.web.app — a separate site, because these builds are AI-written code.