Counted from the logs the agent harnesses write, no estimates. Since 2026-07-13, updated weekly. Last run 2026-10-11 07:17 UTC.
A token here is every token a vendor billed for: prompt input, cache writes, cache reads and output. Cache reads are the cheapest kind and the largest share: an agent re-reads its whole context on every turn. The page says what it could not count, so the number is a floor, not a guess.
| Vendor | Total | Input | Cache read | Cache write | Output | Share |
|---|---|---|---|---|---|---|
| Claude (Claude Code, whole fleet) | 33.81 B | 6.0 M | 32.46 B | 1.21 B | 130.7 M | 55.1% |
| Claude (older daily ledger, totals only) | 17.34 B | 0 | 0 | 0 | 0 | 28.2% |
| OpenAI Codex CLI (hub) | 10.24 B | 457.0 M | 9.75 B | 0 | 34.9 M | 16.7% |
| OpenRouter: GLM, Mistral (hub) | 22.1 M | 18.0 M | 0 | 0 | 4.1 M | 0.0% |
OpenRouter is pay-per-token: 34.17 USD so far. Claude and Codex run on flat subscriptions, so no dollar figure is shown for them.
| Model | Vendor | Tokens | Share |
|---|---|---|---|
| claude-opus-5 | claude | 17.56 B | 28.6% |
| claude (archive ledger, no model split) | claude-archive | 17.34 B | 28.2% |
| gpt-5.6-sol | codex | 7.07 B | 11.5% |
| claude-opus-5-5 | claude | 6.57 B | 10.7% |
| claude-sonnet-5 | claude | 4.39 B | 7.1% |
| gpt-6-astra | codex | 2.92 B | 4.7% |
| claude-fable-5-1 | claude | 2.45 B | 4.0% |
| claude-fable-5 | claude | 1.54 B | 2.5% |
| claude-sonnet-5-5 | claude | 756.1 M | 1.2% |
| claude-opus-4-8 | claude | 351.8 M | 0.6% |
| codex-auto-review | codex | 150.8 M | 0.2% |
| claude-sonnet-4-6 | claude | 142.9 M | 0.2% |
| codex (model not in log) | codex | 55.3 M | 0.1% |
| glm-5.3 | claude | 36.1 M | 0.1% |
Every machine in the fleet uploads its own Claude usage slice once a day. Codex and OpenRouter are counted on the hub only. Machines with a slice: A-2022BAYAREA, FLEET-ANCHOR, MacBook-Anton, MacPro192, ZBOOKG8-2023PAL.
| Machine | Claude tokens | Share |
|---|---|---|
| A-2022BAYAREA | 29.02 B | 56.7% |
| ZBOOKG8-2023PAL | 10.69 B | 20.9% |
| MacBook-Anton | 5.90 B | 11.5% |
| MacPro192 | 4.02 B | 7.9% |
| FLEET-ANCHOR | 1.52 B | 3.0% |
| ISO week (Monday) | Claude | Codex | OpenRouter | Total |
|---|---|---|---|---|
| 2026-W41 2026-10-05 | 17.96 B | 1.20 B | 4.4 M | 19.16 B |
| 2026-W40 2026-09-28 | 11.25 B | 2.01 B | 11.4 M | 13.27 B |
| 2026-W39 2026-09-21 | 4.60 B | 5.20 B | 6.4 M | 9.81 B |
| 2026-W38 2026-09-14 | 1.58 B | 936.3 M | 0 | 2.51 B |
| 2026-W37 2026-09-07 | 5.60 B | 298.6 M | 0 | 5.90 B |
| 2026-W36 2026-08-31 | 9.55 B | 149.4 M | 0 | 9.70 B |
| 2026-W35 2026-08-24 | 611.9 M | 616.7 k | 0 | 612.5 M |
| 2026-W34 2026-08-17 | 0 | 159.7 M | 0 | 159.7 M |
| 2026-W33 2026-08-10 | 0 | 134.8 M | 0 | 134.8 M |
| 2026-W32 2026-08-03 | 0 | 133.6 M | 0 | 133.6 M |
| 2026-W31 2026-07-27 | 0 | 4.7 M | 0 | 4.7 M |
| 2026-W30 2026-07-20 | 0 | 3.8 M | 0 | 3.8 M |
| 2026-W29 2026-07-13 | 0 | 2.1 M | 0 | 2.1 M |
The first and the current week are partial. Weeks older than the 45-day retention of the Claude fleet ledger are frozen from earlier runs and never shrink.
token_usage_ledger.py nightly over its Claude Code transcripts, deduplicates usage records by message id (streaming writes the same record several times), and drops one hourly row per session and model into the fleet bus. This page merges all slices and collapses a machine that reports under two hostnames into one.token_count after every turn. We take the deltas between consecutive events, so a session that spans weeks is split correctly.llm_ask.py writes one line per call with the input and output tokens and the price the API returned.Generated 2026-10-11 07:17 UTC by a weekly job on the fleet hub. Not affiliated with Anthropic, OpenAI or any other vendor named here.