Token Reduction Tools Don't Save Money: Cutting 38% Output Saves Only 1.3% of Bill
JFPuget · x · 2026-09-02
A study testing three token reduction tools against a Claude Code baseline found that even the most aggressive tool, which removed 38.4% of delivered tool output tokens, actually increased the bill by 6.8%.
Key Findings:
- The "harness base" (system prompt + tool definitions) accounts for 74.7% of the bill, which external tools cannot touch.
- Everything a compressor can reach adds up to only 6.0% of the bill, with tool outputs alone being just 3.3%.
- Cutting 38.4% of tool output translates to roughly 1.3% savings on the total bill.
- Cache creation and reads constitute about 80% of the bill.
Conclusion: The amount of tokens a tool cuts tells you nothing about what you actually pay, as the major costs lie in the system prompt and caching.
More from Infra
- NVIDIA Hosts Q&A on NeMo Switchyard via Nemotron Labs — NVIDIAAI · 2026-09-02
- Sequoia invests $100M in Form Energy's iron-air batteries for the AI grid — DavidCahn6 · 2026-09-02
- DuckDB async I/O boosts speed: HF dataset reads 2-3x faster — vanstriendaniel · 2026-09-02
- GitHub Auto-Ban on Core Contributor Sparks Developer Concerns — braelyn_ai · 2026-09-02
- Open-sourced doc parsing pipeline hits 118 pages/min on a single H200 — Low_Acanthisitta7686 · 2026-09-02
- einopx: A unified JAX-native array pattern API — A_K_Nain · 2026-09-02