# Sonnet 5.5 built a 30 s motion-graphics reel. Same as Opus 5.5 quality, but 2x cheaper. — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/3f9697535f958056

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-09-29 13:53 ingested 2026-09-30 02:43

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1wtapu5/sonnet_55_built_a_30_s_motiongraphics_reel_same/)

## retained text



```
Video attached. Everything in it is code: canvas frames drawn in headless Chrome, piped to ffmpeg, with a numpy-synthesized soundtrack. No editor, no stock footage, no video model.

The prompt, in an empty folder, on Sonnet 5.5 with ultracode on:

>make a dynamic 30-second motion graphics video about Reddit that shows what an incredible motion designer you are. and not too fast, people should actually understand what the video is about.

**How the Workflow tool was used for the 30 s cut:**

* 5 builder agents, each owning exactly one scene file, told not to touch the shared files (timing table, compositor, lib). I did those edits myself while they ran.
* 5 independent reviewer agents, told not to edit anything. They rendered contact sheets, read the PNGs, and returned a structured defect list (severity, beat, suggested fix).
* 4 fixer agents ran only where a reviewer said "fix". The fifth shot got "ship" and skipped that stage. That was a `pipeline()` with a condition, so each shot moved on independently.
* 14 agents total, about 45 minutes, 618 tool calls.

**What made it work:**

* **Verification is in the repo.** `shots.mjs` makes contact sheets or full-res frames, `bench.mjs` gives ms per frame. Agents were told to actually Read the images and iterate. That's how a "word never appears" bug and a text overflow got caught without me.
* **File ownership by prompt, not by worktrees.** All agents shared one folder. That only worked because every prompt said which one file was theirs and to use unique screenshot names. I skipped `isolation: worktree` on purpose so they could all render the full project.
* **Reviewers found real bugs**, including an `arc()` with a negative radius that would have killed the final render at one exact frame. It only survived by luck of the motion-blur sub-sample timing.
* **Background jobs + Monitor for the render**, not sleeping. The harness blocks foreground `sleep`, which is a good nudge. 1,800 frames took about 6.5 min.

**Token profile.** Max 20x, so no bill. These are API list-price equivalents from the session transcripts, deduplicated by message id:

* 739k output, 101.6M cache-read, 3.07M cache-write, about 1k fresh input
* Orchestrator: 97 API calls, $11.65. Sub-agents: 442 calls, $23.75. Total about **$35.40** at Sonnet 5.5 list prices ($2 / $10, cache reads $0.20; cache writes assumed 1.25x input)
* The same tokens at Opus 5.5 prices ($4 / $20, cache reads $0.20) would be about **$50.50**. That's a calculation, not a run. 99% of tokens are cache reads, which cost the same on both models, so the gap is about 43%, not 2x.

I did the same kind of prompt on Opus 5.5 earlier and I think the Sonnet cut holds up. It's not a controlled test, though: separate sessions, separate days, and an Opus run would use different token counts.
```



document fetched 2026-09-30 00:44 · canonical https://reddit.com/r/ClaudeCode/comments/1wtapu5/sonnet_55_built_a_30_s_motiongraphics_reel_same
