A 73-second film where every frame is code, built by one Claude Fable 5.2 session
What one long coding-agent session produced, and the numbers it took.
A robot runs out of battery in the middle of a job and goes to the night market to recharge. That is the whole story of "Before Morning". What makes it worth writing down is how it was made.

What was built
- 73.5 seconds of film, 16:9.
- Every frame is code. The scenes are Three.js: geometry, lights and motion written out, rendered frame by frame. No image model and no video model drew anything.
- One session. A single Claude Fable 5.2 session in Claude Code wrote it, checked its own frames and fixed what it saw.
What it took
| Wall-clock time | 3 h 13 min |
| Requests | 286 |
| Output tokens | 462,579 |
These are the session's own counts, not an estimate.
What it cost
We do not state a figure yet. The list price of Claude Fable 5.2 is $8 per 1M input tokens and $40 per 1M output tokens, so the output alone is about $18.50 at list. The larger part of a long agent session is input read back from cache, and how that is billed differs by provider and seller. Until that is confirmed for this run, a total would be a guess.
What can be said: the same session, with the same tools, runs through a gateway by changing one line, the base URL. What that does to the bill is what the price pages on this site track.
Why this matters if you pay for tokens
A long session is where the price per token stops being abstract. 286 requests is an ordinary afternoon for a coding agent. If you run agents on your own card, the two numbers to watch are your output tokens and how much of your input is served from cache.
Sources
- The session's own log for the film "Before Morning" (start and end times, request and token counts)
- Anthropic's announcement of Claude Fable 5.2, 2026-11-09, for the list price