The local machine
43 plates, 61 sprites
About 55 minutes of GPU time produced 299 outputs in all: 107 images, 25 edits for poses and variants, 76 cut-outs, 42 depth maps and 49 ambience and effect clips, from which these made the final cut.
Real-time film · 19:29 · production notes · 8 October 2026
The Long Reach
A nineteen-minute film about three million years of invention, told as a gallery of cut-paper shadow boxes. It was written, illustrated, animated, scored and rendered in one working session by thirteen agents and the studio’s own machine. This is where the time and the money went.
The session in four numbers
From a brief to a published film
5 h 52 minfrom the brief at 06:42 to the published film at 12:34
25 h 12 minof agent working time, run in parallel
13 agentsone Opus orchestrator, six Haiku researchers and six Sonnet animators
$243.04if every call had been billed at API list prices
19:29 film14 chapters61 shots111 lines, 2,766 words4 voices43 plates63 sprites49 ambience and effects13,000 lines of film code6 research notes
The day
06:42 to 12:34
Each bar is one agent, from its first instruction to its last report; each thin mark is a tool call. The four local rows mark every picture, sprite, voice line and sound the machine produced. The six researchers started first, the six animators started together once the engine and plates were ready, and the orchestrator reviewed and built the opening, chapter 1 and the finale in between.
Opus 5.5Sonnet 5.5Haiku 5.5Local machine, no API cost
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 |
|---|---|---|---|---|---|
| Orchestrator · Opus 5.5 | |||||
| Director: orchestration, review and export | 06:42–12:34 | 5 h 52 min | 386 | 496 | $69.43 |
| Research · Haiku 5.5 | |||||
| Research: stone tools, fire, cave hands | 06:51–07:33 | 42 min | 20 | 85 | $0.82 |
| Research: steam, electricity, telegraph, radio | 06:51–07:21 | 30 min | 28 | 109 | $0.67 |
| Research: computing, internet, AI | 06:51–08:01 | 1 h 10 min | 34 | 144 | $0.84 |
| Research: writing, paper, print, compass | 06:51–07:49 | 58 min | 36 | 107 | $0.82 |
| Research: wheel, lever, pulley, machines | 06:51–07:52 | 1 h 01 min | 49 | 90 | $0.71 |
| Research: farming, shelter, pottery, metal | 06:51–07:50 | 59 min | 23 | 99 | $0.75 |
| Animation · Sonnet 5.5 | |||||
| Animate chapters 10-11 (computing, web) | 08:20–09:57 | 1 h 37 min | 230 | 259 | $17.06 |
| Animate chapters 8-9 (steam, electricity) | 08:20–11:43 | 3 h 23 min | 511 | 562 | $45.22 |
| Animate chapter 12 (AI, the mirror) | 08:20–09:55 | 1 h 35 min | 176 | 196 | $14.67 |
| Animate chapters 2-3 (fire, farming) | 08:20–10:55 | 2 h 35 min | 407 | 451 | $31.01 |
| Animate chapters 6-7 (print, navigation) | 08:20–11:00 | 2 h 40 min | 390 | 419 | $32.43 |
| Animate chapters 4-5 (wheel, writing) | 08:20–10:31 | 2 h 11 min | 367 | 398 | $28.60 |
| 13 agents | 2,657 | 3,415 | $243.04 | ||
The bill
Where $243.04 went
By role
Orchestrator · $69.43 · 29%Direction, the paper-theatre engine, chapters 0, 1 and 13, the score and all review.
6 Haiku researchers · $4.61 · 2%One research note each, from stone tools to AI.
6 Sonnet animators · $168.99 · 70%Two chapters each, and chapter 12 on its own.
By model
By kind of token
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 50 min in all.
The animators
How long each pair of chapters took
Chapters 8 and 9, steam and electricity, ran longest and set the pace for everything that needed all six.
Six animators, two chapters each
The film
What the day produced
Narration
111 lines
Kokoro read 2,766 words in a few minutes, 107 lines by the narrator and the quotations in three other voices, between 08:12 and 08:16.
The score
One score, many ages
Written in code to grow through the ages, from breath, drum and bone flute to lyre, keyboard, machine ostinato, Morse rhythm, square-wave arpeggios and choir, with one hand motif in every age.
The MP4
35,062 frames
Drawn by a headless browser on the GPU in 32 minutes, from 12:01 to 12:33, and encoded to an 888 MB H.264 file. A scan of 2,338 frames, one every half second, found no errors.
What the checks caught
Where the animators’ hours went
The first report split each animator’s time between checking and writing code. Checking meant rendering frames, reading contact sheets and fixing what they showed.
| Animator | Working time | Where it went |
|---|---|---|
| chapters 2-3 (fire, farming) | 2 h 35 min | 65% checking, 30% code |
| chapters 4-5 (wheel, writing) | 2 h 11 min | 63% checking, 28% code |
| chapters 6-7 (print, navigation) | 2 h 40 min | 67% checking, 28% code |
| chapters 8-9 (steam, electricity) | 3 h 23 min | 67% checking, 22% code |
| chapters 10-11 (computing, web) | 1 h 37 min | 50% checking, 44% code |
| chapter 12 (AI, the mirror) | 1 h 35 min | 43% checking, 44% code |
Choosing models
The same tokens on another model
The identical requests priced at the other model’s rates. A different model would not have used exactly the same tokens, so these are comparisons, not forecasts.
| Work | As run | The same tokens on another model |
|---|---|---|
| Animation | Sonnet 5.5: $168.99 | Opus 5.5: $337.98 |
| Research | Haiku 5.5: $4.61 | Sonnet 5.5: $19.82; Opus 5.5: $39.65 |
What it taught
Five things the numbers say
The animators were the bill
The six Sonnet animators cost $168.99, 70% of the total, almost all of it re-reading their own conversations of 780,000 to 910,000 tokens at their largest. Full-size contact sheets and large source files built those conversations up. One chapter per agent, a fresh agent for each round of fixes and half-size review images would plausibly cut that.
Checking is where the polish came from
About two thirds of most animators’ time went on rendering, looking and fixing, and about 30% on writing code. That loop is what makes the animation hold up, and it is also what fills the conversations.
The model choices paid off
Research on Haiku cost $4.61; the same tokens on Sonnet would have cost $19.82. Animation on Sonnet cost $168.99, half of the $337.98 the same tokens would have cost on Opus.
The slowest agent set the pace
Chapters 8-9 (steam, electricity) took 3 h 23 min and chapter 12 (AI, the mirror) took 1 h 35 min. The orchestrator spent 3 h 02 min in gaps of more than 20 minutes between events, mostly waiting for reports, and the first report estimated that splitting the heavy chapters more evenly could have saved about an hour and a half.
The studio machine did the heavy lifting
Every image, cut-out, depth map, sound, voice and video frame was made locally, and none of it appears on the bill.
Method
How these figures were measured
Measured from the session's own transcripts: the director's, and one for each of the 12 agents it launched. Assistant messages are deduplicated by message id and the last entry's usage is kept. Models come from each agent's metadata. Figures run to 12:34, when the film was published. This page was rebuilt with the shared script and gives a total of $243.04; the first report, built with an earlier version of it, said $243.92.
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, Sonnet 5.5 at $2 and $10, Haiku 5.5 at $0.10 and $0.50 (higher rates above a 100,000-token prompt), 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). Local-machine marks are the modification times of the generated files. The 299 outputs and the GPU time come from the first report’s count of ComfyUI’s output files, and the render, scan and narration figures from the same report. The hour-by-hour split sorts tool calls by keyword, and gaps of more than 20 minutes with no events are left out of it.