RTK Token Savings Don't Cut Coding Bills, Benchmark Finds

RTK (Rust Token Killer) compresses terminal output before your coding agent reads it — 79k GitHub stars, and a viral X post claiming 60% fewer Claude Code tokens at 313K views. Today Quesma published the first serious cost audit: $1,500 spent, 1,740 attempts on Terminal-Bench 2.1, five runs per task with and without RTK. The receipts don't match the marketing.

What the benchmark found

With RTK, Claude Code with Fable 5.0 spent 5% less total; OpenCode with DeepSeek V4 Pro 0813 spent 5% more. On a per-task basis — the fairer weighting — Fable was 1% more expensive (statistically zero) and DeepSeek tasks cost 17% more. Fable's entire saving traced to a single task, corewars; exclude it and savings were under 1%. The "89% reduction" RTK reported across 349.2M claimed saved tokens? Two head -1 calls against a big file credited 120.5M tokens each — output those commands would never have returned. rtk gain measures removed bytes, not billed tokens, and it can make a more expensive attempt look optimized.

Why the math collapses

Three forces cancel the compression. First, terminal output is a small slice: ~11% of Fable's input tokens, ~40% of DeepSeek's — and Read/Grep/Glob tools bypass RTK entirely. Half of Claude Code's Bash calls already self-limited with head/wc. Second, agents cache: rereads of terminal output bill at 1/10 (Fable) to 1/30 (DeepSeek) of regular input, so the raw-token denominator is inflated. Third — the killer — extra turns. DeepSeek's RTK runs took 18% more turns at 7% less input each; 58 tasks took more turns and 44 of them cost more. One RTK incompatibility looped an agent into 339 consecutive errors, 9× baseline cost. JetBrains's independent SkillsBench run saw the same pattern. Frontier models already conserve tokens themselves — this is a niche optimization being sold as a bill-cutter.

Verdict

Don't reach for RTK to save money. The lesson is bigger: any tool that reports "savings" in raw bytes while your bill is computed in cached input, extra turns, and model reasoning is a marketing counter, not a billing counter. Ask for dollars, not bytes.