One Opus run cost $1.99 — 94% went to reading files, not writing code
aakashgupta · x · 2026-11-11
The author broke down a real Claude Opus 5.5 run over a large open-source Java codebase: 1.09M tokens, $1.99 total, with only $0.13 (6%) spent on writing the answer. By dollars, cache writes (first read of each big file at $8/M tokens) were 85% of the bill; cheap cache re-reads were 80% of tokens but just 9% of cost. The finding validates Spotify's trick of handing file reading to a cheaper model, claimed to cut token use 90%, which the author then tested on two codebases.
More from coding & agent
- Dev fine-tunes Qwen3-1.7B into a 1.83GB GLB 3D asset coder that runs on CPU — ConfidentDinner6648 · 2026-11-11
- Solo dev shares 20k-word workflow: Opus 5.5 + Seedance 2.5 for 4 viral product videos — dotey · 2026-11-11
- CodeGraph turns codebases into Neo4j graphs so agents can do dependency-aware refactoring — JeremyCMorgan · 2026-11-11
- Primitive imagines guardian agents that enforce team decisions infrastructure-wide — daniel_mac8 · 2026-11-11
- Fable 5.2 builds a visual workspace for watching Claude sub-agents work — FinanceYF5 · 2026-11-11
- Fable 5.2's 7 early experiments: 3D cells, verified origami, agent visualization — FinanceYF5 · 2026-11-11