Numbers only, measured from my own local Claude Code session logs (2026-10-12 to 2026-10-18, 7-day daily average).
- Calls per day: 2,772
- Cache-read tokens per day: 315M
- Output tokens per day: 2.4M
Share of cost by slice:
| Slice | Share of cost |
|---|---|
| Base (system prompt + tool list) | 43.3% |
| Model replies | 25.5% |
| User input + attachments | 21.6% |
| Shell command output | 6.1% |
| Subagents | 2.1% |
| File reads | 1.2% |
| Search (grep/glob) | 0.1% |
| MCP + other tools | 0.1% |
| Web fetch | 0.0% |
Cost weights: input 1, cache write 1.25, cache read 0.1, output 5 (input token = 1).
Takeaway: the biggest slice is the fixed context that gets re-read on every call, so shorter sessions help more than trimming tool output.
Top comments (2)
Interesting split :) One thing I'd check in your weights - in my logs every main-session cache write was the 1-hour kind, which is billed at 2x input, not 1.25x. If yours are the same, that moves a fair chunk of money into writes. In 33 days of my logs priced at API rates, cache writes were 29% of cost and output only 8%. The expensive writes were cold restarts. 207 times I came back to a session after the hour was up and the whole context got written again, $972 or 22% of main-session spend (against $48 if it had been warm). So shorter sessions help, but only if you END them rather than leave them to go cold. Moral of the story dont go for lunch or sleep, otherwise Anthropic will bleed you dry 😂
Cold restarts are the cost a token total hides. Once the cache window closes, the whole context is written again in full, so a session left to go cold can cost more than the work in it.
If you want to see where your own sessions leak, the THOL benchmark (public, raw runs) measures session cost end to end: pi-infected.github.io/token-harnes... . On long sessions (over 200k tokens) with Claude Code, Tokenade came out 39% lower on average (measured 2026-09-14). That is an average, not a ceiling.
I built Tokenade, so weigh it as such.