The local machine
106 images
Plates and pencil-drawing edits from 06:48 to 08:55: 52 base paintings and 54 edits, generated at a resolution of 1440 (1920 by 1088). The soundtrack is your own song, so nothing was voiced or scored.
Real-time film · 6:44 · production notes · 19 October 2026
What the Rain Made
A six-minute, forty-four-second music video for your song The Celtic Fire, painted in white paper and blue watercolour by the weather itself: rain lands on the page, a pencil drawing appears under it and blue blooms from each drop, from Loch Lomond to the stones, the sea and the Saltire the rain made. One Fable 5.1 director and one Haiku researcher made it in a morning. This is where the time and the money went.
The session in four numbers
From a song to a finished music video
3 h 00 minfrom the brief at 06:26 to the commit at 09:26, with about an hour of that waiting on you or the GPU
3 h 03 minof agent working time
2 agentsone Fable 5.1 director and one Haiku 5.5 researcher
$27.69if every call had been billed at API list prices
6:44 film1920×1080 at 3063 shots106 generated images0 narrated lines1 song, your own570 lines of film code3,200 words of research
The morning
06:26 to 09:26
Each bar is one agent, from its first instruction to its last report; each thin mark is a tool call. The row below marks every picture the machine produced. The gaps in the director’s bar, about an hour in all, are times with no activity, while the GPU worked or the agent waited for you.
Fable 5.1Sonnet 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 |
|---|---|---|---|---|---|
| Director · Fable 5.1 | |||||
| Director: orchestration, review and export | 06:26–09:26 | 3 h 00 min | 139 | 224 | $27.58 |
| Research · Haiku 5.5 | |||||
| Research Scotland facts for film | 07:10–07:13 | 3 min | 13 | 66 | $0.11 |
| 2 agents | 152 | 290 | $27.69 | ||
The bill
Where $27.69 went
By role
Director · $27.58 · 99.6%Concept, pilot, 63-shot storyboard, 106 image prompts, engine, review, export and the commit.
Researcher · $0.11 · 0.4%One fact sheet of about 3,200 words.
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. 1 h 59 min in all.
The film
What the morning produced
The rain engine
570 lines
A shared pool of blooming masks, one scene per shot with special handlers, a per-shot table and measured anchors for the table, hearth, window and fork paths.
The timing authority
A song map
Cuts and blooms are keyed to song-map.md, copied from the third Celtic Fire version. The part-two turn at 2:08 is a single drop that falls upward, and the slam at 5:41 floods the Saltire.
The MP4
404 s, 473 MB
1920 by 1080 at 30 fps, CRF 20 with no denoise, about 9 Mbit/s, with an 81.6 MB single-file page. A 195 MB share copy at 4 Mbit/s keeps the paper tooth and pencil in a check of the St Kilda cliff.
What the checks caught
Eleven findings from the pilot and the rough cut
The review log records each diagnosis with the frames inspected, from the four pilot versions to the first full render.
| Found | What went wrong | How it was fixed |
|---|---|---|
| Pilot v1: the rain edit | Came out as night, with chorus frames nearly black-blue and the paper gone; blooms read as dark bubbles | Reprompted as a soft shower in pale tones with white paper kept; blooms redrawn as crinkled blots |
| Pilot v1: rings and streaks | Water rings were invisible on blue and rain streaks too faint | Rings as pale screen ellipses with a thin dark inner line; streaks wider and darker, then slanted into the wind |
| Pilot v2: the rim | A rim band far too wide and dark, a crackle pattern over the hills | Narrowed to 0.965 to 1.03 of the blot and faded with the age of the bloom |
| Pilot v3: the rim | The rim now reads as a real wet edge and the 40 s picture is clean; the edits were still at the default size | Regenerated at a resolution of 1440 (1920 by 1088), and build.py given a hung-job watchdog |
| First full render | A page error: the scene context was passed to drawImage instead of its canvas, which the Node smoke stub accepts | Fixed so every handler draws the scene canvas |
| Glencoe stag | The stag was literally red in the first plate | Reprompted with a russet-brown coat and a new seed |
| Letter shot | A tan desk and tan paper, so the whole frame left white paper | Reprompted on white writing paper and a pale desk |
| The slam | The Saltire’s blue was near navy and dirty from the dark mottle | A cleaner base blue with a lighter mottle |
| Gallery pull-back | It ended at 1x, so the 54 cards were too small | Ends at 1.4x |
| Luskentyre | Used twice within 30 seconds | The second shot became the St Kilda stac with gannets, painted upward |
| Table plate | The empty-table edit kept the tartan rug, so the rug appears before its item blooms | Left: it reads as part of the chair; minor |
The GPU
Two times the card looked hung
You noticed high utilisation with a cool card twice, and asked the director to check.
| When | What you saw | What it was |
|---|---|---|
| 07:50 | GPU utilisation high but temperature low | A stall. The director added a five-minute watchdog to build.py that restarts the server and retries a hung job |
| 08:30 | The same symptoms again | Not a hang: the puffins’ pencil edit had just landed, and the card was pulling 426 W at 71 °C on the next job |
What it taught
Six things the numbers say
A cheap film from an expensive model
The Fable 5.1 director cost $27.58, 99.6% of the $27.69 total, because Fable 5.1 lists at $10 per million input tokens and $50 per million output. The film still came to $4.11 per finished minute, among the cheapest measured, because the whole film was one agent and a few small helpers.
Your pilot feedback reshaped the film
You liked the look and said a whole minute on a single scene was too long. The film became 63 shots, 106 generated images and fast cuts from small subjects to big ones, with several items painted into one scene.
Image generation was the main work
The director spent 27 min on plates and images, and 36 min on export and publishing, against 12 min on code, because the rain engine is small and the pictures carry the film.
The GPU stalled once and was working the second time
Two of your messages flagged high utilisation with low temperature. The first was a real stall, which led to a watchdog. The second was a long job, which the director confirmed from the power draw and the log.
Stills, not playback, were inspected
The rough cut was reviewed from 189 frames, three per shot, read at half size. The log says playback cadence was not inspected, and nobody has yet reported on the film’s playback or sound.
Finished and committed in one session
The session ended with a 195 MB share copy at 4 Mbit/s with a light denoise and a local commit,
a335ee0onceltic-fire-brief. Nothing was pushed.
Method
How these figures were measured
Measured from the session's own transcripts: the director's, and one for each of the 1 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. The director ran on claude-fable-5-1, priced from the published rates: $10 per million input tokens, $12.50 for five-minute cache writes, $20 for one-hour writes, $0.25 for cache reads and $50 for output. Figures run to the commit at 09:26, and the transcript contains three stretches of activity separated by gaps of more than 20 minutes.
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 pilot and rough-cut findings come from the project’s review log, the GPU notes from the director’s messages, and the share copy and commit from its final 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.