Our own numbers
We run our own agent team on our efficiency tools, including the JidoSeal model. Here is what happened to its work per token, week by week in 2026. These are our own team's numbers. Your results will differ. They are real and measured, and they show a direction, not a guarantee.
The table
| Week of | Tasks | Context tokens read (billions) | Tasks per billion | Cost per task | Tasks per $100 | Turns per task |
|---|---|---|---|---|---|---|
| 08-31 | 78 | 0.41 | 191 | $4.1 | 24.4 | 50 |
| 09-07 | 67 | 0.70 | 96 | $8.3 | 12.0 | 75 |
| 09-14 | 70 | 0.79 | 89 | $15.0 | 6.7 | 78 |
| 09-21 | 84 | 1.11 | 76 | $16.1 | 6.2 | 77 |
| 09-28 | 93 | 3.06 | 30 | $34.4 | 2.9 | 116 |
| 10-05 | 116 | 2.27 | 51 | $21.2 | 4.7 | 100 |
| 10-12 | 80 | 1.26 | 64 | $19.8 | 5.1 | 74 |
| 10-19 | 121 | 1.23 | 98 | $14.2 | 7.0 | 63 |
| 10-26 | 123 | 0.56 | 219 | $6.0 | 16.7 | 49 |
How we measured. Claude agent tasks that finished, not counting runs made through the API, grouped by the Monday of each week, at list price. The source is our own task logs. We rechecked the count with a second run on 2026-11-01. The week of 10-26 was still open at that reading.
What the table shows
About 7x more tasks per billion tokens from our worst week to our latest: 30 in the week of 09-28, 219 in the week of 10-26. Cost per finished task went from $34.40 to $6.00 across the same two weeks, at list price.
The sturdier figure. Context read per turn fell about 53%: from 186,876 before our context cap (768 tasks) to 87,258 after it (122 tasks). It fell on the same model too: Opus 5.5 from 237,514 to 108,122, and Sonnet 5.5 from 104,114 to 67,503 (first reading, 119 tasks).
How strong this is
Real and measured, and directional.
- "About 7x" runs from the worst week to the latest week. It is not exponential. The first week on record was already at 191, so most of it is a recovery to the August level.
- Job size fell too. Median cost per task went from $12.52 in the week of 09-28 to $0.83 after the cap. So 7x overstates a like-for-like gain.
- Our scorecard is more modest: tasks per $1,000 at list price went 44.5, 54.3, 57.0, 80.1 from the week of 09-28 to the week of 10-19.
What else changed
Several things changed at the same time:
- The context cap, live on 10-26.
- Routing work to Grok first, from 10-22.
- Tool deferral, from 10-19. Tool schemas went from 127,238 to 54,847 characters.
- Lighter models, with Sonnet and Haiku on 77% of tasks.
- Weekly-cap throttling.
A replay estimated the cap alone at 50% fewer reads and 36% less net input. That is an estimate, not a measurement.
What this says about OKF
OKF's share is not isolated. We added Gold frontmatter to 625 files from 08-05 to 08-26. That was before our first complete token log, so there is no baseline. Work per token also fell through September with Gold already in place. So the honest sentence is this: these results come from our efficiency tools, including the JidoSeal model.
Design-partner results
None yet. We'll add them here as they land.
Pick a folder, scan it, and see where your own files stand. It works with Claude Code, Copilot, Cursor, Codex, Gemini CLI, Windsurf and local models. Scan a folder, free.