Across 3,171 sessions and 30B tokens, only 0.3% was model output — a local tool that searches your agent logs
Rare_Guide_9830 · reddit · 2026-09-14
A developer noticed Claude Code and Codex write every session to disk as JSONL (per-response token counts, tool calls and results, model used, compaction points), so he built a local tool that parses them into a searchable SQLite DB with a UI.
His year's data:
- 3,171 sessions, 30.1B tokens recorded — but only 92.7M generated output: 0.3%
- 112,849 model responses; 97.3% of input served from cache
- 0.3% of tokens ≠ 0.3% of cost: output bills at roughly 50x cached input, so volume and money split very differently
Why: every turn resends the conversation until compaction; output stays flat while input compounds — agents read far more than they write and pay to re-read it every turn.
Unexpected find: 3,171 tool-call failures across 116k calls he'd never noticed; some created usage-draining loops, 95% inside sub-agent processes.
More from coding & agent
- WebMCP browser tools cut tokens 52% on one task but increase them on another, DeepDeck experiment finds — j032 · 2026-09-14
- DeepDeck launches WebMCP tool directory with inspectable source and pinned revisions — j032 · 2026-09-14
- Agora shows how to build voice AI apps with your coding agent — Embarrassed-Bend7110 · 2026-09-14
- Claude Code + Video CLI Pitfall: Server Errors While Jobs Still Run Led to Double Charges — letandrewcook · 2026-09-14
- Desktop Commander plugin turns ChatGPT Pro into a local coding agent via your terminal — RileyRalmuto · 2026-09-14
- Webagent: Open-Source Harness Turns Any Website Into a Talking Agent in Minutes — Scobleizer · 2026-09-14