Claude Code tokens per day, per session, per turn: measured
One machine, one heavy user, 2026-07-20 to 2026-09-23: 36 active days, 160 sessions, 11,145 assistant turns. "Context tokens" = input + cache read + cache creation, the quantity plans meter.
Daily distribution
| Per active day | Median | p75 | p90 | Max | Mean |
|---|---|---|---|---|---|
| Context tokens | 14,220,380 | 95,874,785 | 272,969,392 | 662,268,182 | 84,384,384 |
| Output tokens | 92,683 | — | 1,347,841 | 2,234,869 | — |
Per session and per turn
| Measure | Median | p90 | Max |
|---|---|---|---|
| Context tokens per session | 3,204,396 | 14,762,676 | 1,138,969,622 |
| Assistant turns per session | 34 | 91 | — |
| Output tokens per session | 26,589 | — | — |
| Context tokens per turn (n = 11,145) | 185,177 | 625,517 | 997,880 |
The per-turn maximum sits just under 1,000,000: that is the context window filling up. Sessions near it pay the most per turn, which is why the 8,000-character cap matters only for a fifth of sessions and matters a lot there.
Ten heaviest days
| Date | Context tokens | Turns |
|---|---|---|
| 2026-07-22 | 662,268,182 | 1,425 |
| 2026-08-11 | 468,049,303 | 1,553 |
| 2026-08-10 | 320,705,743 | 2,430 |
| 2026-07-24 | 284,273,571 | 412 |
| 2026-07-21 | 261,665,212 | 511 |
| 2026-07-23 | 182,527,496 | 460 |
| 2026-08-05 | 156,059,940 | 667 |
| 2026-09-23 | 129,529,285 | 388 |
| 2026-08-31 | 115,382,120 | 716 |
| 2026-08-17 | 89,372,340 | 515 |
Last 21 active days
| Date | Context | Output | Turns | Sessions |
|---|---|---|---|---|
| 2026-08-19 | 48,345,892 | 171,007 | 170 | 4 |
| 2026-08-20 | 77,426 | 268 | 1 | 1 |
| 2026-08-21 | 80,932 | 268 | 1 | 1 |
| 2026-08-24 | 0 | 0 | 1 | 1 |
| 2026-08-25 | 0 | 0 | 5 | 5 |
| 2026-08-26 | 0 | 0 | 3 | 3 |
| 2026-08-27 | 0 | 0 | 11 | 11 |
| 2026-08-28 | 0 | 0 | 16 | 16 |
| 2026-08-29 | 0 | 0 | 11 | 11 |
| 2026-08-30 | 0 | 0 | 10 | 10 |
| 2026-08-31 | 115,382,120 | 1,017,477 | 716 | 10 |
| 2026-09-01 | 50,692,374 | 271,535 | 330 | 5 |
| 2026-09-02 | 48,133,515 | 264,839 | 300 | 6 |
| 2026-09-03 | 61,170 | 250 | 1 | 1 |
| 2026-09-04 | 64,971 | 250 | 1 | 1 |
| 2026-09-05 | 68,834 | 250 | 1 | 1 |
| 2026-09-06 | 0 | 0 | 2 | 2 |
| 2026-09-07 | 0 | 0 | 1 | 1 |
| 2026-09-13 | 0 | 0 | 1 | 1 |
| 2026-09-19 | 0 | 0 | 1 | 1 |
| 2026-09-23 | 129,529,285 | 1,590,290 | 388 | 4 |
Models seen
claude-opus-5: 6,160 turns · claude-opus-4-8: 2,438 turns · claude-fable-5: 2,075 turns · claude-fable-5-1: 338 turns · <synthetic>: 73 turns · claude-opus-5-5: 49 turns · claude-haiku-4-5-20251001: 12 turns
The kit: measure your own sessions and the market, weekly
- Session cost anatomy for your own logs (
quota_report.py): the exact report on this site, on your machine, one command; reads only usage fields and lengths. - Demand simulator (
demand_sim.py+ priors + calibration): live connectors for GitHub, npm, VS Code Marketplace, WordPress.org, Hacker News, Apify, MCP registry, Hugging Face; scores what a solo newcomer reached at each attach-point in the last 24 months, calibrated against six real product outcomes before it scores anything. - This week's datasets: every table on this site as JSON and TSV, the plugin marketplace catalog, the issue gap maps, the 50-candidate scoring run with every verdict and its deciding number.
$39, one time. Regenerated weekly for a year; the download link you get keeps serving the newest build. MIT.
Get the kit, $39Card checkout by Stripe on this site's own account; the download page opens the moment payment completes. Questions: riley@agentexchange.work.
Pulled 2026-09-23 · usage fields of assistant messages in ~/.claude/projects/*/*.jsonl on one machine (one user); content never read