97% of coding-agent tokens are cache reads, but cache writes eat 31% of the cost
Nitjsefnie · reddit · 2026-11-14
A Claude Code user priced every session transcript since January at API list prices — about 205B tokens — and split costs by token type:
- Cache reads: 97.2% of tokens, 58.0% of cost
- Cache writes: 2.5% of tokens but 30.6% of cost, with 1h TTL writes alone at 23.7% (1h writes cost 2x input, 5m writes 1.25x)
- Output: 0.3% of tokens, 10.9% of cost; fresh input: 0.07% of tokens, 0.6% of cost
Two pricing gotchas: cache-read multipliers are no longer uniform — Opus/Sonnet 5.5 are 0.05x while Fable 5.1/5.2 are 0.025x, so a flat multiplier misprices them; and Claude Code logs one streamed reply multiple times under the same requestId, so sum lines and you double-count — keep the largest.
The author open-sourced claudit (MIT), a self-hosted dashboard that reads Claude Code, Codex and Kimi CLI transcripts from your own bucket, with side-by-side token/cost panels and a read-only live demo.
More from coding & agent
- WordPress UTM QA tip: accept cookies and submit forms without refreshing — HaktanSuren · 2026-11-14
- Using Claude to Visualize Ideas in Real Time During Team Calls — Flomerboy · 2026-11-14
- Vercel Token Pass doubles value: $150 in open-weight model tokens for $30/month — cramforce · 2026-11-14
- Dev moves nearly all code changes to subthreads, parent thread just orchestrates — zeeg · 2026-11-14
- Code-generated 3D Where's Waldo with 10,000+ animated people playable in browser — CSProfKGD · 2026-11-14
- Coding agent Junior goes macro-aware, sharing context across parallel sessions — zeeg · 2026-11-14