What these tokens would have cost billed per token at each model's published price. Not a bill. It is the one figure that compares across people on different plans.
List-rate estimate / day
Jun 17, 2026peak $477.31Aug 4, 2026
on a plan$8,752
paid per token$130.36
per active day$181.27
9.17B
tokens
187.2M on a working day
59,448
requests
154.3K each
49
active days
not a day missed
49
day streak
still running
Activity
9.17B tokens over 49 active days · last 12 months
Averaging 187.2M on a day worked, with a median of 169.4M. Busiest was Jul 29, 2026 at 885.5M. Hover a day for the models and harnesses behind it.
lessmore
Last 12 months in full on this profile, last synced Aug 4, 2026. Counters only: no prompts, no project names, no file paths.
Drift
What got reached for, month by month
Each row is one month at 100% of its own tokens, so a quiet month reads the same width as a loud one and the only thing changing is the mix.
Output is a fraction of the tokens and usually most of the money, because it is priced at several times the rate of anything read back out of cache.
cache read
$4,402
output
$2,782
cache write
$1,512
fresh input
$186.43
Every model in this archive has a known rate. 3% of them ran on models whose id we can only pattern-match, so their rate is our reading of a public price rather than an exact match.
Cache
Reuse, not volume
A raw token total mostly measures how much context got re-read. The ratios underneath it are the part that says anything about how the work was done.
98.9%
of everything entering the context window came from cache instead of being paid for again
91.2×
tokens of context read per token written
154.3K
tokens in an average request
cache read
96% · 8.78B
cache write
2% · 202.4M
output
1% · 99.5M
fresh input
1% · 96.2M
Models
36 models, mostly claude-opus-4-8
claude-opus-4-8
2.38B · 26%
$2,578 estimated at list rates · 15,468 requests
claude-opus-5
2.16B · 23.6%
$1,526 estimated at list rates · 8,695 requests
claude-fable-5
1.64B · 17.9%
$3,548 estimated at list rates · 9,435 requests
claude-sonnet-5
1.17B · 12.8%
$458.26 estimated at list rates · 10,343 requests
cursor-grok-4.5-high-fast
550.4M · 6%
$253.13 estimated at list rates · 862 requests
grok-4.5
328.6M · 3.6%
$127.14 estimated at list rates · 2,201 requests
cursor-router-auto
275.9M · 3%
$26.03 estimated at list rates · 390 requests
claude-sonnet-4-6
150.6M · 1.6%
$139.04 estimated at list rates · 2,235 requests
claude-haiku-4-5
123.7M · 1.3%
$35.33 estimated at list rates · 3,326 requests
gemini-3.6-flash
86.8M · 0.9%
$28.95 estimated at list rates · 662 requests
+26 more below one percent.
Tools
10 harnesses, one archive
Claude Code
7.67B · 53,901 requests · $6,677
83.6%
Cursor
914.2M · 1,352 requests · $319.74
10%
Grok Build
328.7M · 2,208 requests · $127.20
3.6%
Antigravity
102.9M · 917 requests · $40.09
1.1%
OpenCode
92.8M · 792 requests · $2.19
1%
Claude Code rollup
58.7M · 92 requests · $1,711
0.6%
Codex
4.0M · 95 requests · $4.69
0%
Cline
386.7K · 79 requests · $0.14
0%
OpenClaw
230.1K · 11 requests · $0.05
0%
Copilot CLI
17.0K · 1 requests
0%
Rhythm
Heaviest on Monday
Tokens by day of the week, in the timezone each machine reported.
Mon: 1.65B over 7 days
Tue: 1.06B over 7 days
Wed: 1.35B over 7 days
Thu: 1.64B over 7 days
Fri: 996.0M over 7 days
Sat: 1.31B over 7 days
Sun: 1.16B over 7 days
MTWTFSS
By month
3 months tracked, busiest Jul 26.
Jun 26: 727.2M over 14 days
Jul 26: 7.62B over 31 days
Aug 26: 829.5M over 4 days
JunJulAug
Hour of day is not in the sync payload yet, so the rhythm here stops at the day. See docs/tokenstat-sync-cli.md for what the CLI would need to send.
Spread
Across 63 projects
Project names stay on the machine. What reaches us is an opaque hash we can count but never read, so a profile can show how the work was spread and never which work it was.
project 1
32.3%
project 2
26.2%
project 3
10%
project 4
10%
project 5
8.3%
project 6
4.7%
project 7
2.9%
project 8
2.1%
Ordered by size, unlabelled on purpose. The top project is 32.3% of everything.
How the counts were read
Some harnesses report token counts exactly; others have to be derived from what they write to disk. The archive keeps that distinction rather than flattening it, so a number that was inferred never passes itself off as a measurement.
