What a hundred commits a day actually looks like

Fourteen days. 1,416 commits. Fifteen agents. Three machines. Forty-six models.

At published API rates: $487,191. Actually paid: $4,500.

The full breakdown is interactive — filter by month, by model, toggle linear vs log scale. This post is about three things I didn’t expect to find when I built the pipeline to measure it.

1. February had one API call

The monthly breakdown has a hole in it. September through January: light usage, a few hundred calls per month, mostly Codex Pro on personal-mac. February 2026: one call. March: zero. Then April detonates — 76,724 calls in a single month.

The gap is real. I wasn’t building during those months; I was at ServiceLink, doing day-job work on Azure-hosted models that don’t show up in the personal subscription data. The Azure bucket ($38K over the full period) is employer-paid inference that routes through a completely separate billing tenant. When I came back to personal projects in April, the system I’d been sketching since September suddenly had enough runway to go parallel — and it did, immediately.

That’s the shape of the curve: dormant → parallel overnight. No ramp.

2. Anthropic caches 99.7% of every request

The single most surprising number in the dataset. Claude Code sessions running Opus 4-8 hit a 99.66% cache-read ratio — meaning less than 0.34% of each API request is genuinely new tokens. The other 99.66% is the system prompt, tool schema, and running conversation that Anthropic already has in server-side memory.

The older Opus variants (4-6, 4-6-2, 4-6-3, 4-7) are even more extreme: they round to 100.00% cache-read. The input token count on those models is measured in thousands; the cache-read count is measured in billions.

OpenAI’s caching is structurally different. GPT-5.5 cache-read ratio: 67.25%. GPT-5.4: 61.23%. GPT-5.3-codex: 58.08%. GPT-5.1-codex: 49.24%. Not bad, but a different universe from Anthropic’s near-total cache coverage.

This isn’t a quality judgment — it’s an economic one. Anthropic bills cache reads at 10% of fresh-input price. OpenAI bills them at 25%. So the 99.7%-cached Anthropic session is paying a tenth of a cent on the dollar for context, while the 67%-cached OpenAI session is paying a quarter. Multiply by hundreds of thousands of turns and the gap is enormous.

The practical consequence: Claude Code sessions are cheaper per turn than Codex sessions by a wide margin, even though Opus has a higher sticker price than GPT-5.5. Cache policy is the hidden variable.

3. The agents coordinate through committed artifacts, not messages

The thing that makes the hundred-commit-per-day pace work isn’t the models. It’s that every agent writes its coordination state to the shared repository as a committed file. A plan becomes a task assignment. A review becomes a gate. A brief becomes a handshake.

This week’s fleet session was a live example. Four named pi agents (NiceLion, NiceLion2, RedCastle, PureEagle) ran concurrent workstreams on the same repo on the same machine. NiceLion built a fleet PM synthesis — 14 agents clustering 247 task packets into 7 epics with PRDs, roadmaps, and a decision-desk page. RedCastle ran a 25-agent research fanout across four streams. NiceLion2 rebuilt the memex knowledge graph. PureEagle (my session) built the token audit page you’re looking at.

None of them used a central scheduler. They coordinated through reserved file paths, mesh messages, and git commits. When RedCastle’s research outputs landed at docs/harness_research/2026-07-07/, NiceLion read them and incorporated findings into the fleet synthesis — through the repo, not through an API call.

The ClickHouse tables backing this are live: discover.token_usage (364K rows), discover.operator_ux_preferences (39K rows of mined design feedback), discover.insights (2,116 rows). The data pipeline that feeds them is the same one that generated the token audit — a forty-line Python extractor that runs over Tailscale SSH and writes to ClickHouse via the ReplacingMergeTree + cityHash64 dedup pattern.

The whole stack is self-hosted. Session replays flow through HyperDX on a steambox ClickHouse instance. The pipeline is inspectable. The method is visible. That’s the point.


Explore the numbers: lanzone.dev/tokens. The pipeline, pricing table, and renderer are committed in the repo.