Benchmarks
Measured, not promised.
Every number tokentrim quotes, with how it was measured and on how much data. Measurements are tagged measured; anything projected is tagged estimate.
History re-read per setting
A replay of real transcripts: each session is re-run through tokentrim's rules and compaction point, and the context tokens sent on every call are summed before and after. Re-measured 20 Oct 2026.
| Setup | Compacts at | Cut | |
|---|---|---|---|
| Output trimming only | — | 6.9% | measured |
| balanced | 300k | 62.1% | measured |
| strong (default) | 200k | 71.1% | measured |
| max | 150k | 74.9% | measured |
Most of the cut comes from compacting sooner, not from trimming. Trimming tool output alone is under 7%, because Claude Code already keeps big outputs out of the context.
Browser and sub-agent rules
Newer rules, measured on smaller samples. Treat these as directional.
| Rule | Effect | Sample | |
|---|---|---|---|
Screenshot cap 0.4 (screenshot_scale_cap, opt-in) | −20% of screenshot tokens | replay of one week's sessions, 331k screenshot tokens | measured |
lean sub-agent vs general-purpose | −86% tokens | same task run once each | measured |
lean-browser sub-agent for 3+ step browser jobs | ≈ −6% of main-chat input | simulation on recent sessions | estimate |
Drop earlier screenshots in a batch (batch_last_shot_only, off by default) | ≈ −3.5% of tool-result weight | simulation on recent sessions | estimate |
What no tool can cut
Claude Code's own system prompt and tool list cost 40–60k tokens on every call, and a fresh desktop chat starts near 77k. Those are fixed by Claude Code, so a short session saves little and nobody can honestly promise 90% across all your work.
Billed cost and limit burn fall less than tokens, about 55–70%, because cached re-reads are cheap per token and output isn't touched.
Get your own number
The replay runs locally on your own transcripts before you install anything. Nothing leaves your machine.
git clone https://github.com/suncal/tokentrim && cd tokentrim bin/tokentrim replay
Method and code: README · The honest numbers and replay.py.