The model writes a program, not a video

Here is the fact that makes everything else make sense: Claude Opus 5.5 takes text and images in and puts text out. It cannot emit an MP4. So every one of those "AI-made" motion videos is really a program the model wrote, which something else turns into frames.

The engine of the whole trend is determinism. Rather than animate against a live clock, the model writes one function that paints the exact frame for any moment in time. We will call it seek(t): give it a time in seconds, it draws that frame. A headless browser calls seek(t) hundreds of times — 900 times for a 15-second clip at 60fps — screenshots each frame, and ffmpeg stitches them into a video.

Why determinism matters so much:

• Reproducible. No timers, no Math.random, no state carried between frames. Frame 812 is the same every single run. That means you can render just the seconds you changed, and a tweak is a one-line edit plus a re-render — not a re-roll of a black-box video model.
• Seekable. Because any frame is a pure function of t, you can jump straight to the fast moment at 4.2s to inspect it, or render a contact sheet of one frame per beat without simulating everything before it.
• Ownable. The output is code you can read, diff, and reuse. A video model gives you pixels you cannot edit; this gives you a program you can.

This is not a fringe trick. It now has a name — "video as code" — and real tooling. HeyGen open-sourced HyperFrames (Apache-2.0) to do exactly this from HTML; React developers use Remotion to do it from components. Both capture a deterministic program frame-by-frame in headless Chrome and encode with ffmpeg. Different front doors, same idea underneath.

One honest note on why Opus 5.5 is good at this: Anthropic's own launch described a test where several models built a game from one prompt, and Opus 5.5 scored highest on the strength of its graphics and polish. Writing a clean, animated visual program is squarely in its wheelhouse. But the model is still only writing code — which means your job, for the rest of this course, is to build the machine that runs and improves that code.