Code-Compression Bench

Context-compression layers for coding agents, measured like for like.

Every layer run behind the identical headless Claude Code scaffold on 100 tasks from SWE-bench Verified, priced cache-aware and graded by the official Docker harness. One agent, one model, one grader — only the compression layer changes, so any difference in cost or quality is attributable to it.

Run 2026-07-04 · 100 tasks from SWE-bench Verified · model claude-sonnet-4-6 · ranked by cost per solved task

Leaderboard

Ranked by cost per solved task — cache-aware total cost ÷ Docker-graded solves. Click any column header to sort.

Parsec Baseline Woz Headroom RTK
# Arm Solved $/solved Total $ vs base $ Input (M) vs base in Wall h vs base time Steps Peak ctx Cache R:W

Every arm runs the same model through the same scaffold; only the compression layer differs.

Compare any metric

Pick a KPI to compare across arms; the bars re-sort highest to lowest.

Savings vs solve rate

Cost savings versus the no-compression baseline (up) against solve rate (right); the upper-right region is cheaper and solves more. Hover any point.

Cost vs time

Cost savings against time savings, both versus the no-compression baseline; the upper-right region saves on both. Hover any point.

Fact vs fiction

What each layer advertises, next to what we measured behind a real coding agent. None of these headline numbers is measured on the cache-aware dollar cost of solving real repository tasks; they are measured on single-shot QA, shell output in isolation, or favorable-case "up to" ceilings.

LayerTheir headline claimWhat we observed (n=100)
ParsecCurates the agent’s working context−39% cost, −54% input, 62/100 solved
Woz"Cut your Claude Code costs in half"; "30–40% faster" (an estimate)−13% cost, −35% input, 55/100 solved; wall-clock slower, not faster
RTK"60–90% fewer tokens on common dev commands"+13% cost, +16% input, 54/100 solved; its hook rewrites only shell output — native Read/Grep/Glob bypass it
Headroom"60–95% fewer tokens, same answers"; "better caching"+44% cost, +6% input, 58/100 solved; its docs: "code passes through" uncompressed
Caveman"Cuts 65% of output tokens (measured)"; "full technical accuracy"−19% cost, −19% input, 58/100 solved; the 65% is output-only — a small share of agentic-loop cost

Full sourced breakdown — every quote, primary source, the exact benchmark each number was measured on, and the mechanism behind each gap: FACT-VS-FICTION.md.

Method: success = the official SWE-bench Verified Docker grader (fail-to-pass resolved, pass-to-pass intact; no LLM judge). Cost = cache-aware Sonnet-4.6 pricing applied identically to every arm's real per-call usage (uncached input $3.00 / cache-write $3.75 / cache-read $0.30 / output $15.00 per 1M). Full data: paired.csv · summary.json · benchmark sponsored & operated by Dasein, publicly reproducible.