The renderer
Raw WebGL2
A hand-written hidden-line renderer: ink edges, red-chalk construction lines, a bistre wash and white-chalk light on toned paper. Every frame is a pure function of time, so it seeks, pauses and exports frame by frame.
Real-time film · 2:51 · production notes · 1 October 2026
Promethean: the West, in line
A two-minute, fifty-one-second film about what the West gave the world, drawn as architectural models on a staircase of plinths and carried by one flame. It was written, drawn, scored and published as a playable WebGL page in a little over two hours by seven Opus agents. This is where the time and the money went.
The session in four numbers
From a list of beats to a published film
2 h 08 minfrom the brief at 20:09 to the published page at 22:16 on 1 October
8 h 47 minof agent working time, run in parallel
7 agentsone director, one researcher and five vignette builders, all Opus 5.5
$145.04if every call had been billed at API list prices
2:51 film25 drawn models70 bars at 100 BPM9,168 lines of code4,893 words of research notes0 generated images0 samples in the score
The evening
20:09 to 22:16
The director opened at 20:09 and started the researcher at 20:29. At 20:51 it launched all five builders together, one per era, and spent the rest of the evening on the renderer, the score and the timeline while they worked.
Opus 5.5, all seven agentsEach bar runs from an agent’s first instruction to its last report; each thin mark is a tool call.
The company
Who did what, for how long, at what price
| Agent and assignment | Ran (local) | Time | API calls | Tool calls | Cost at list prices |
|---|---|---|---|---|---|
| Director · Opus 5.5 | |||||
| Director: orchestration, review and export | 20:09–22:16 | 2 h 08 min | 209 | 277 | $42.97 |
| Research · Opus 5.5 | |||||
| Research facts and model dimensions | 20:29–21:13 | 43 min | 131 | 226 | $9.48 |
| Vignette builders · Opus 5.5 | |||||
| Build antiquity vignettes (A) | 20:51–21:53 | 1 h 02 min | 92 | 122 | $16.42 |
| Build early-modern vignettes (C) | 20:51–22:16 | 1 h 25 min | 87 | 100 | $21.06 |
| Build machine-age vignettes (D) | 20:51–21:53 | 1 h 02 min | 112 | 135 | $14.45 |
| Build medieval-Renaissance vignettes (B) | 20:51–22:08 | 1 h 17 min | 121 | 138 | $19.78 |
| Build modern-age vignettes (E) | 20:51–22:02 | 1 h 11 min | 156 | 173 | $20.89 |
| 7 agents | 908 | 1,171 | $145.04 | ||
The bill
Where $145.04 went
By role
Five vignette builders · $92.59 · 64%The 25 drawn models, built in parallel from one specification file.
Director · $42.97 · 30%Direction, the renderer, the score, the timeline, review and publishing.
One researcher · $9.48 · 7%Dimensions and facts for every model.
By kind of token
Everything ran on one model, so the lever was context size, not model choice.
The director, hour by hour
Each gap between two transcript events is given to the tool call or message that ended it, then sorted by tool name, file path and command. The buckets are keyword matches, so read them as proportions. 2 h 07 min in all.
The film
What the evening produced
The score
All from code
Synthesised with no samples at 100 BPM, 70 bars, growing from a drone and felt piano to the Moon-launch climax and a last statement of the theme in D major. Picture, camera and text sit on the same bar grid.
The checks
Four lines changed
The facts were checked at the end. Antikythera became “a computer out of bronze”, the Athens label became 5th–3rd century BC, the atom label became Berlin 1938 and Chicago 1942, and the cathedral became Notre-Dame.
The MP4
Made later
The session ended with the playable page. The 1080p, 30 fps MP4 (5,148 frames, 171.6 s, 112 MB, plus a 48 MB share copy) was rendered on 9 October, outside this session.
What it taught
Four things the numbers say
The builders were the bill
The five vignette builders cost $92.59 of $145.04, or 64%, and each one worked inside a context of 580,000 to 700,000 tokens. Almost all of that was cache reads: the same long conversation read back at every step.
One specification file let five agents work at once
VIGNETTES.md gave every model its size, its anchors and its timing in bars local to its own section, so the builders started together at 20:51. Two finished by 21:53, and the slowest, the early-modern group, ended at 22:16, which is when the page could be published.
Everything in code made it fast, and set its limit
With no generated pictures and no narration the GPU never ran, and the film cost $50.83 of list-price API time per finished minute. The remake that followed spent $13.05 per minute on a film eight times as long.
The score was checked by measurement only
The director could not listen, so it checked the loudness arc, the spectral balance, clicks and timing against the picture, and said so in its hand-over. Whether it sounds right has always been a question for a listener.
Method
How these figures were measured
Measured from the session's own transcripts: the director's, and one for each of the six agents it launched. Assistant messages are deduplicated by message id and the last entry's usage is kept, since the first entry of a message can hold a fifth of its final output tokens. All seven agents ran on Opus 5.5 (the model field of every agent transcript says so).
Cost is what the usage would have cost at API list prices in October 2026: Opus 5.5 at $4 per million input tokens and $20 per million output, cache reads at 0.05 times input, cache writes at 1.25 times (five minutes) or 2 times (one hour). A subscription session was not billed this way, so treat the dollar figures as a measure of work, not an invoice.
Times are UK local time (BST, UTC+1). Agent working time is the sum of each agent's first-to-last transcript event; the session ran 128 minutes, and a stray later event on 10 October is left out. The hour-by-hour split sorts tool calls by keyword.