Claude Code token usage: why caveman didn't cut my bill

Описание: Claude Code token usage with the caveman plugin: output fell 23.1% over 5 sessions, it added 3,887 tokens to every API call, and the session bill at API list price did not measurably change.

caveman's README says its skill cuts 65% of output tokens. I ran the same three-turn Claude Code session 5 times with the caveman plugin and 5 times without, on a small Python repo with three failing tests: fix the tests, explain the bugs for a pull request, review for edge cases. Claude Code 2.1.270, Opus 5 at its defaults, 14-15 Sep 2026. Every session fixed the bugs and passed the tests.

WHAT IT DID, PER TURN (medians)
Fix turn, tool calls and code: 30.4% more expensive with caveman, in all 5 sessions. Almost all of it is cache: 13.4c to 18.5c, because the plugin's 3,887 tokens are written to the cache at double the input price and re-sent on every call. Output added 0.7c.
Review turn, a long answer: 24.9% cheaper, 4 of 5 caveman sessions under every normal one.
Whole session: $0.413 normal against $0.443 caveman at the median (+3.4% summed over all 5). Identical sessions spread $0.046 normal and $0.099 caveman, wider than the $0.030 gap. No measurable change.

WHAT I DID NOT TEST
Longer sessions, other models or effort levels, caveman's separate proxy, rtk, or how a Pro or Max subscription counts these tokens. Dollars are API list price as Claude Code reports them, and no money was spent. JetBrains ran a much larger test on Sonnet 5 and measured output tokens down 8.5%.

RAW FILES: https://github.com/parseai-yt/coding-...

CHAPTERS
0:00 The claim
0:21 What a session sends
0:43 How a turn is priced
1:03 What caveman adds
1:30 What caveman cut
1:48 It depends on the turn
2:18 The whole bill
2:39 When it pays
3:06 What I did not test
3:21 Recap

SOURCES
caveman README and docs/HONEST-NUMBERS.md, commit 15581d1, read 14 Sep 2026: https://github.com/JuliusBrussee/caveman
JetBrains AI blog, caveman A/B, July 2026, read 14 Sep 2026: https://blog.jetbrains.com/ai/2026/07...
Claude pricing, read 14 Sep 2026: https://platform.claude.com/docs/en/a...

Send this to whoever on your team installed caveman to save money.

--
ParseAI. I run the thing before you ship it, and show the raw output.
The field ships techniques faster than anyone evaluates them, so most advice is untested.
Adopt fewer things, with reasons.

#claudecode #aiengineering #tokenusage