Research · snapshot 2026-10-01
What we learned cataloguing 4,077 AI motion works.
Claude Opus 5.5 shipped on September 22, 2026. Within days, people were posting motion graphics it wrote as code and rendered to video. We inventoried 128 public GitHub repos (works came from 92 of them) and the posts they point to, then de-duplicated them. Here is what the corpus says.
- Unique works indexed
- 5,454
- Attributed to Opus 5.5
- 4,077
- Creator-published prompts
- 1,195
- Distinct prompt families
- 981
- Share their code
- 12%
- Vertical (9:16)
- 12%
1. Prompts travel. Results don't.
The most reused prompt in the corpus is a one-line brief asking Claude for a 15-second showreel. 109 separate works came from it, and no two look alike. See them side by side. A prompt gets you something like the video you saw; only the source gets you that video.
2. Almost nobody ships the code.
Only 12% of Opus 5.5 works link to their source, and just 105 works publish a prompt, a video and code together. The prompt is what people ask for first (“what's the prompt?” is the reflex reply under viral posts), but the code is what makes a clip reusable.
3. Regenerating is expensive and slow.
Creators who shared their numbers typically reported $2–15 of API spend per clip, with outliers from $25 to $100, and anywhere from 10 minutes to 7 hours of iteration. That cost repeats every time someone tries to reproduce a clip from its prompt.
4. The tools are open, and agent-ready.
Most work names no framework at all, which usually means a single HTML file rendered with a headless browser. Where it's named, HyperFrames, Remotion, Three.js, shaders and GSAP lead. Several ship official skills and llms.txt files for agents. Browse all 17 tools.
5. Vertical is scarce.
Only 12% of works with a known aspect ratio are 9:16, despite where short video is watched. That gap is an opportunity for creators.
Method and caveats
Works were de-duplicated by X post id, media id, video URL and repository. Model attribution is what each creator reported. Framework and category labels are keyword-based and unverified per item. Engagement figures came from third-party snapshots. The site lists all 4,323 works. 2,474 have a still or video, 1,053 a prompt and 395 an open licence; the rest show a title, a credit and a link to the original, and are kept out of search engines until their creator adds more.