The tokens, counted
Every token this studio has spent was already on disk in two places nobody had read. Thirteen working days, 6,163 messages from a machine, all measured rather than estimated.
7,195,750 output tokens. The single biggest day was 1 September at 2,427,539, which is thirty-four percent of the whole project, and that was the day eleven synthetic seats were hired.
The raw total including cache traffic is 2,562,098,421, and 97.9% of that is cache-read: the same context re-read at every turn. It grows with the length of a conversation, not with anything made, and it bills at roughly a tenth of input price. Publishing "2.5 billion tokens" would say this studio is three hundred times larger than it is.
So output leads and the breakdown sits under it. It is the only line that tracks work produced rather than conversation length.
What one night costs
Anchoring on the allowance consumed this week, one percent is about 1,047,370 output tokens.
| Output tokens | Share of a week | |
|---|---|---|
| Night 003, whole run | 170,727 | 0.16% |
| Five doubled nights, Monday to Friday | 1,707,270 | 1.6% |
A full working week of doubled nights costs 1.6% of the allowance. The nights are not what consumes it. The daytime work is.
The limits that do bind are three, and none is tokens. Wall clock, of which night 003 used 7.9%. Turns, which were capped at 400 and are now 800. And context length, which is the one with evidence behind it: the Chief of Staff measured quality degrading at six seats in one run, because the specifications read at the start were far back in memory by the sixth seat, so the run copied the previous run's output instead of the spec.
That decides how the headroom gets spent. More nights and more separate seat calls, not one longer session.
The hat
An image was wanted for the first article: a gunfighter looking straight at the camera. Three attempts each produced the same four faults, including a hat silhouette from 1865 that the studio's own list bans, and a background from a John Ford film that three prompts asked it not to draw.
The list gets a concession. Technology stays inside the era, because the game is built on it: black powder, fire your shots and then close, revolvers as expensive status weapons rather than ordinary kit. Clothing does not. A hat from the wrong decade in a fictional borderland costs nothing a reader notices, and enforcing it costs attempts that produce nothing else.
Then a style add-on fixed what the prompts could not. Same seed, same words, one variable changed:
| Three prompt-only attempts | With the style model | |
|---|---|---|
| Medium | A digital render every time | Oil paint, brushwork, canvas |
| Hat | The banned silhouette, three for three | Absent |
| Background | Monument Valley, three for three | The flat prairie that was asked for |
| Weapon | Present despite "no weapon visible" | Absent |
It was chosen over better-known ones because it was trained entirely on public-domain images. This project's reference painters are out of copyright on purpose.
The placeholder that nearly went out
The article was scheduled to send at 20:00. It carried a bracketed note reading "CHART GOES HERE", left in because the diagram tool was not installed and installing it was tonight's job.
Installing that tool to remove a dependency exposed it, seven hours before the send. The chart renders correctly and is in the article now: two solid lines where a working relationship has carried real work, ten dotted ones where it has only been written down.
Found by checking a precondition, not by a reader finding it.
Noah Revoy CEO, Lone Hand Studio