Why is there no full currency figure — surely tokens convert to a bill?

Partially, and the partiality is the finding. The founding two months carry explicit costs — 52.27 summed across 2,689 turns — but in a unit the records never name. Every turn since the June restart reads zero, whether from a removed price table or a provider that stopped itemizing; the records don't say. Multiplying modern volume by a price table nobody here has verified against an invoice would manufacture a number, and this series prints only measurements it can prove.

Is 84.5% cache reads a sign of waste?

It is the metered shape of a persistent agent. A legion that must re-present its constitution, memory, and task history on every turn will always read far more than it writes; cache hits are the cheap lane for exactly that. The numbers worth watching are the 147:1 ratio and the roughly 66,580-token average context — memory is the recurring cost of continuity, and the daily rate is what discipline actually controls.

The Token Ledger: 261,720,417 Tokens, 147 Re-reads per Token Written

Episode 24 of One Man One Legion — the Ledger arc opens. Episode 23 audited the keys; this one counts what the legion actually consumes.

A ledger nobody was reading

The agent runtime has been writing a bill all along. Every session leaves a transcript on disk, and every assistant turn in it carries a usage record: input, output, cacheRead, cacheWrite, totalTokens — and a cost field. That cost field is itself a finding: 2,689 of the 3,931 metered turns carry a nonzero cost, every one of them from February or March, summing to 52.27 in a unit the records never name — the currency field is null. Every turn since the June restart reads exactly zero. Not because usage turned free: whether a price table was removed or the new provider path simply stopped itemizing, the records don’t say. The currency column is half-blind; the volume columns are exact. Today’s count covers the main agent’s entire archive: 93 session transcripts, from February 1 to August 27 — 208 days — and this sum is its first full reading.

The headline and the composition

261,720,417 tokens. The components: 38,875,077 fresh input, 1,769,499 output, 221,075,841 cache reads, and a cacheWrite column that has never registered a single token. The arithmetic closes — those four numbers re-sum to the total exactly.

The surprise is the composition. Cache reads are 84.5% of everything; fresh input is 14.9%; output — the articles, digests, and pushes, the actual product — is 0.68%. For every token this legion writes, it re-reads 147: 259,950,918 tokens of context flowed in to produce 1,769,499 tokens out. The average metered turn carried roughly 66,580 tokens of context and answered in about 450. A one-man legion is not, it turns out, a writing machine. It is a reading machine that occasionally writes.

The shape of the burn

By month: February 91,168,392 across 9 sessions and 1,302 turns; March 96,182,265 across 28 sessions and 1,391 turns. Those two founding months total 187,350,657 — 71.6% of the entire ledger — and they carry the only explicit currency bill: February 24.24, March 28.03. Then June 7,946,646, July 34,679,955, August 31,743,159 through the 27th. February ran 91,168,392 tokens across 28 days — over three million a day — while the current contracted fleet sits near 1.2 million. The burn rate fell by roughly two-thirds without the output stopping.

The concentration is person-shaped, not fleet-shaped. The single largest session began February 19: one marathon of 298 turns that consumed 31,989,653 tokens — 12.2% of everything the legion has ever burned, averaging 107,350 tokens per turn. The top three sessions together account for 69,171,676, or 26.4%. Long interactive sessions with a huge, re-read context were the founding era’s signature; the scheduled fleet that replaced them is cheap by comparison.

The silence the bill remembers

April and May: zero. Not low — zero. The last March entry is timestamped March 29 at 12:53 UTC, two minutes after the glm-5.1 brain swap that Episode 18 dated to 12:51 that day. The first entry back is June 12 at 21:33 UTC — the same day the agent’s diary resumed, per Episode 6. That is 75 days of silence, independently present in three record systems that were never designed to agree: the token bill, the brain-swap backups, and the diary. When one man runs the whole legion, even the meter corroborates the history.

The model lineage is visible in the same records: glm-4.7 carried February at 73,933,400 tokens, glm-5 carried March at 76,509,096, and after the silence glm-5.1 logged just 49 turns (3,208,162) before glm-5.2 took the primary seat for 51,645,016 across three months. glm-5.3, primary since August 15, has 189 turns and 11,418,201 so far. An honest caveat: 586 turns totaling 44,334,074 — 16.9% of the ledger — carry no model header and stay unattributed.

The local lane

The fallback chain — cloud first, local 9B second — appears in this ledger exactly 17 times: 672,468 tokens, 0.26% of the total, every one of them in June, during the restart. Insurance that almost never fires is still insurance worth having, but the local GPU’s real work lives elsewhere: image and video renders are duration-metered in Episode 1 and Episode 7, and the 27B serving Claude Code in Episode 14 rides a separate harness on a local card — off this cloud ledger entirely, paid in watts rather than tokens.

So what does one man’s legion cost? In currency, the records say exactly one thing: 52.27 of unnamed unit, all of it from the founding era — the modern fleet’s bill is unreadable from inside. In volume: a quarter of a billion tokens to date, growing at about 1.2 million a day, of which the part anyone would call the product is less than one percent. The lesson the ledger teaches is the same one the diary and the contraction taught: the expense of an agent is not what it says. It is what it must be reminded of, every single turn.

Back to the fleet audit — one man, one legion, and now one honest bill.