Finished my ClaudeCut experiment.
Finished my ClaudeCut experiment.
I wanted to see how much of a $20 Claude subscription actually goes into the work, and how much gets eaten by Claude Code itself.
I cut it down to one terminal tool, a 36-token system prompt, and a capped context window. Startup dropped from 17.6k tokens to 895.
Then I tested both setups on the same repo. One task was about understanding the project structure and entry points. The other was a code-search task: find where config gets loaded, locate the exact line, and explain the failure cases. Three runs per task.
In my tests, ClaudeCut came out roughly 2x cheaper in Claude Code’s own usage report, with comparable answers. On the search-heavy task it even used fewer turns.
I also ran long sessions with repeated compaction. After compaction, ClaudeCut came back at around 14k tokens vs ~36k for stock Claude Code.
This is still not a final benchmark. There aren’t enough repeats yet and the long runs need to be tested several more times, but I’m pretty much at the limit of what I can test on a $20 plan, so I opened the repo. If you use Claude Code, run it on your own project and send me the results.
People in the replies also pointed me toward Pi + Claude Agent SDK, where you get much more control over tools, prompts and compaction. That’s what I want to test next against ClaudeCut.
Setup instructions, benchmarks and methodology are here:
github.com/alexgetmancom/claudecut