Lab token usage, 2026
What a year of agentic coding at home actually consumed. Measured from Claude Code transcripts, cross-checked against the subscription's own meters. Snapshot as of 11 September 2026.
Where the tokens actually go is the argument for building it this way. This is the measurement.
The figures
- ~24B tokens processed (input + output, including cache), 28 May – 11 September
- ~89M tokens generated (model output) over the same window
- Six weeks at the plan's weekly ceiling for the review-class model, 11 August – 15 September
- $11–13k/month API list-price equivalent at September's pace, on assumed prices
Two volume conventions exist and both are true, so quote both:
- Processed is the usual headline. Platforms and usage tools count input plus output, cached context included.
- Generated is the figure for work produced — throughput, and where output cost lives.
~24B tokens processed and ~89M generated by agents since late May 2026. Most of that volume is long-context re-reads, which is where agent cost actually lives.
At least 85% of the API-equivalent cost is cache reads and writes rather than output. August generated 41M output tokens, which comes to at most ~$1k of its ~$7–8k.
By month
| month | API calls | output | cache write | cache read | API-equivalent | basis |
|---|---|---|---|---|---|---|
| Jan–Apr | — | — | — | — | ≈ $0 | no transcripts, no co-authored commits, no PRs |
| May | 1,110 | 2.0M | 6M | 0.10B | $139–160 | a few surviving sessions (floor) |
| Jun | 1,847 | 2.6M | 30M | 0.67B | $586–698 | partial; older transcripts pruned |
| Jul | 20,278 | 16.2M | 171M | 6.27B | $3,665–4,190 | transcripts from ~12 Jul only; full month ≈ $5–8k |
| Aug | 39,718 | 40.9M | 315M | 10.32B | $6,843–7,895 | full month; the most reliable row |
| Sep 1–11 | 20,277 | 27.1M | 127M | 5.68B | $4,236–4,703 | ≈ $11–13k at this pace |
One machine — the laptop I drive the agents from — carries 98% of it. The compute box adds about $214–228 in June, mostly cache reads. The always-on mini adds $6–7 in May.
Every month before August is a floor, not a measurement. Claude Code deletes old transcripts on a default retention period, so the early numbers are what survived.
Cross-check: transcripts against the subscription meters
Two independent meters record the subscription's utilization, and they agree to the percent: on the same three dates, both read 75, 73 and 80.
| quota week | output | cache read | API-equivalent | weekly peak | review-model peak | $ per 1% |
|---|---|---|---|---|---|---|
| 14 → 21 Jul | 0.5M | 0.06B | $59–68 | 37% | 64% | $2 |
| 21 → 28 Jul | 4.4M | 1.22B | $733–825 | 82% | 97% | $9 |
| 28 Jul → 4 Aug | 9.9M | 4.94B | $2,589–2,996 | 99% | 100% | $28 |
| 4 → 11 Aug | 12.6M | 3.27B | $2,044–2,333 | 100% | 92% | $22 |
| 11 → 18 Aug | 9.8M | 2.11B | $1,434–1,648 | 75% | 100% | $21 |
| 18 → 25 Aug | 8.1M | 2.09B | $1,352–1,564 | 73% | 100% | $20 |
| 25 Aug → 1 Sep | 7.5M | 1.66B | $1,327–1,536 | 75% | 100% | $19 |
| 1 → 8 Sep | 18.0M | 3.56B | $2,713–3,014 | 80% | 100% | $36 |
| 8 → 15 Sep (partial) | 8.9M | 2.02B | $1,456–1,615 | 57% | 67% | $27 |
The transcripts account for what the subscription meters from August onward. API-equivalent per 1% of weekly quota holds a steady $19–36 band across six weeks, and no week shows high utilization with low tokens.
July has usage the transcripts do not explain: 14–28 July reads 37% and 82% on the meter but only $2–9 per 1%. That fits heavy chat use in the spring tapering off. It is an inference, not a measurement, and it is the one claim here I cannot stand behind with data.
The binding limit is not the all-model quota. It is the quota for the review-class model, which has sat at 100% every week since 11 August.
Cross-check: what came out
| month | pull requests opened | co-authored commits |
|---|---|---|
| Jan–May | 0 | 0 |
| Jun | 29 | 35 |
| Jul | 252 | 270 |
| Aug | 374 | 272 |
| Sep 1–11 | 396 | 525 |
Same shape as the token data: nothing before June, a step in July, September running about 3× August's rate. Cost per pull request fell from ~$20 in August to ~$11 in September, when many were small documentation and plan changes carrying one review each.
These counts are floors too — the credential that read them sees 35 repositories, and commits were counted across 59 local clones by co-author trailer, de-duplicated by hash.
Method, and what I am unsure about
Source. The per-message usage block in the Claude Code transcripts — input, output, cache write, cache read, with model and timestamp. It includes subagents, headless review runs (437 in September) and the fleet orchestrators, because all of them bill the same subscription.
Counting. De-duplicate by message id and keep the maximum of each field. 18,003 of 81,587 messages record output that grows across streamed lines, and the last line is always the largest. Keeping the first line instead undercounts output by 21% — 68.9M against 87.2M. I found that because two scans of the same data disagreed, which is the only reason I trust the number now.
Prices are assumptions, in USD per million tokens: $5 in / $25 out for the largest models, $3/$15 mid, $1/$5 small; cache read 0.1× input; cache write 1.25–2× input, which is where the ranges come from. List prices for the current model generation were not published at the time of measurement.
Not counted: web chat sessions, which leave no local record, and one service's separately-billed API traffic.
The machinery these numbers came from: Plan dispatch, drawn.