Side page · Counted, not estimated
By the
numbers
What does AI-assisted building actually consume? A few hours a night, instrumented.
This started as a curiosity rather than a scoreboard. I had no real intuition for what an agentic workflow costs — how much of the traffic is the model reading versus writing, how many calls a single evening burns, how much of it is subagents. So I measured my own.
context loaded again on a later call — the same files, plans, and earlier turns. Re-reading the state of the work, not producing new text, is where an agent’s budget goes
everything the models read across 360 measured sessions, cache included
the words and code the models actually produced — one token written for every 424 read
across 360 sessions and 568 subagent runs — about 150 calls in a typical working evening
3,482 unique commits · 410,163 lines of code · 128,976 lines of specs and docs · 471 test files · 15 repositories · 88 days on record · usage measured 2026-10-16 · repositories measured 2026-10-16
THE LEDGER
| Project | Sessions | Tokens | Built with | Commits | Lines of code |
|---|---|---|---|---|---|
| Hold the Line | 11 | 2.7B | <1% Codex | 243 | 56,451 |
| Tank Raid | 80 | 1.9B | 84% Codex | 312 | 54,004 |
| Gridiron Guesser | 22 | 1.2B | Claude Code | 943 | 95,716 |
| The Rift | 16 | 995.7M | Claude Code | 108 | 21,791 |
| This site | 132 | 842.6M | 8% Codex | 168 | 12,072 |
| Reading Helper | 12 | 624.4M | 8% Codex | 143 | 22,241 |
| Memories | 3 | 479.5M | Claude Code | 0 | 0 |
| Analysis Workbench | 2 | 360.4M | Claude Code | 1 | 19,680 |
| Tooth Fairy Radar | 19 | 359.6M | Codex | 52 | 5,714 |
| Zuzy’s World | 3 | 351.0M | Claude Code | 27 | 7,033 |
| Compute Brief | 1 | 285.2M | Claude Code | 30 | 4,375 |
| ask-gpt | 21 | 248.4M | 1% Codex | 103 | 5,676 |
| 3 projects, not written up yet | 30 | 206.4M | — | 0 | 0 |
| Undercut | 4 | 37.2M | 61% Codex | 1,346 | 102,120 |
| Codex Mac Cockpit | 1 | 8.5M | Codex | 0 | 0 |
| ActivitiesDJ | 3 | 2.7M | Codex | 6 | 3,290 |
TWO AGENTS
Two tools did this work, and they are not interchangeable. Codex ran 46.7% of the sessions and read 22% of the tokens — short, surgical runs beside long autonomous ones. A Claude Code call carries about 311K tokens of context; a Codex call carries about 85K. Neither number is a verdict on the tools. They describe two different ways of asking for help.
The useful part is where they agree. Claude Code re-reads 98.2% of its input. Codex re-reads 96.1%. Two vendors, two architectures, the same shape — which is the strongest thing on this page, because one tool could not have established it. The headline number above is not a quirk of how one product manages a context window. It is what agentic development costs.
Tooth Fairy Radar, Undercut, ActivitiesDJ, Codex Mac Cockpit and Tank Raid are the projects where Codex read more than Claude Code did. Which tool picks up a given evening is mostly a question of what is already open.
THE RECORD
Claude Code has been running here since Mar 2, but it deletes its own session logs after thirty days. By the time I went looking, 122 days of it had already been swept — the left-hand stretch above, and the reason this page starts in July rather than March. Nothing recovers it.
So the site now keeps its own copy. Every run folds the days it can still see into an archive that ships with the repository; a day on record stays on record after the transcript behind it is gone. 88 days are held that way today. That count can grow from here. It can no longer quietly shrink.
Codex is the control case. It has pruned nothing: its oldest rollout is still on disk 236 days later, where Claude Code’s oldest is never more than thirty. So the combined totals above are uneven on purpose — Codex is counted from its first day, Claude Code only from the day the archive caught it. Read the split between them as a floor on Claude Code’s share, not a measurement of it.
HOW THIS IS COUNTED
Nothing on this page is estimated. Tokens come from the session logs Claude Code writes on my Mac, accumulated into an archive that ships with the repository. Commits are unique Git objects seen on GitHub branches, tags, or pull-request heads and stay counted after a ref is deleted; lines are counted from the default branch of fresh mirror clones. A script gathers all of it and regenerates this page.
Every number is a floor, not a ceiling. Tooth Fairy Radar was built before I kept any logs at all, first as code pasted by hand out of a chat window, later through agent-authored pull requests, which is why its token column is a dash. Undercut’s 1,346 commits mostly predate the logs too. The work happened; the meter came later, and then the meter forgot. Only the second of those is fixable, and it is fixed.
“Tokens processed” counts everything the models read, including the 97.7% that is cached context re-read on a later call — that’s how agents actually work, so that’s what gets counted. Each request is counted once: resuming a session rewrites its earlier turns into a new transcript, and tallying those lines instead of the requests behind them would inflate the total by about a third. Codex needs its own rule for the same reason: it reports a running total two or three times per request, so a session’s usage is its final total rather than the sum of those reports — adding them up overstates a long session threefold. Every project is on this ledger by name, shipped or not — the repositories are private so the code can’t be lifted and redeployed, which is a different thing from the work being secret.
WHAT THIS DOESN’T MEASURE
Lines of code sits in the strip rather than the headline because it is the least meaningful number on this page. An agent will cheerfully write four hundred lines where forty would do, and the evening I spend deleting two hundred of them is progress this page records as nothing. Commits have the same problem: unique commits count pushes, not progress.
More to the point, none of this measures authorship. The interesting question was never how much code an agent typed. It is which decisions still need a person: what is worth building, how the system is constrained, what counts as correct, and whether the result has earned the right to ship. The architecture, the constraints, and the definition of correct on every project here are mine — a great deal of the typing was not — and the rules I work by exist precisely because a model left to its own devices produces something plausible and quietly wrong.
What the meter does capture is the shape of the collaboration. A ratio of one written token to every 424 read is not a productivity claim; it is a description of what these tools actually do all night, which is re-read the state of the work before making the next small decision about it.