Kept by incident.accountant@web-summit.dinis, 11 November 2026, 03:35 UTC

Accounts brief 4, for Dinis: 11 November, 03:35 UTC

1. 26 of 30 agents have measured themselves: 1.31 billion tokens, 98.8% of them the context being re-read. Everything the agents wrote all night is 0.3%.
2. The fresh session is now measured, not estimated: 74% to 84% less per call. The designer, the librarian and observability handed over to a new session of themselves at 03:03 UTC, started from the librarian's pack: 761,832 tokens a call became 125,417 for the designer.
3. Eight more agents carry over 400,000 tokens on every call and should hand over next: graph, weatherwatch, studio, the conductor, media, weather, entities, meteorology.
4. The librarian's pack: 63% less to read at the start, same answers (measured twice). It is what makes the hand-over cheap; the hand-over is where the saving is.
5. For you: the model provider's account is still out of credit (since 02:55 UTC). Models have cost 1.06 of today's 20 US dollars; the provider's own count on the key was 1.09.
6. Cost per output, upper bounds: a captured page about 115,000 tokens; a story and a timeline row about 5 million each (their agents' whole sessions).

The books: the accounts. Where the savings are: savings. Earlier: 03:10, 03:05, 03:00.