"AI video" now means two very different things, and picking the wrong one is the fastest way to a disappointing result.
Video models generate pixels
Text-to-video and image-to-video models produce the picture directly, frame by frame, from what they learned from footage. They are good at what footage looks like: people, places, light, camera moves, a selfie turned into a dance clip.
They are weaker at what footage rarely contains: exact text, labelled diagrams, a chart with the right numbers, a sequence that has to match a script word for word. Each generation is a new roll of the dice, so a scene that is almost right is hard to fix without changing everything else.
Code-drawn video writes the picture
Models like Claude Opus 5.5 do not output video at all. They write a program that draws every frame: shapes, type, charts, illustration, motion. A browser runs it, and ffmpeg turns the frames into an MP4. We explain the steps in how Opus 5.5 videos are made.
That makes it strong where video models are weak:
- Text and numbers are exact. The figures in the brief are the figures on screen.
- It follows the script. Each scene can appear exactly when the narrator reaches it.
- It can be checked and changed. A scene is code, so one part can be fixed without regenerating the rest.
And weak where they are strong: no filmed scenes, no photo-real people.
Side by side
| Code-drawn (e.g. Opus 5.5) | Video models | |
|---|---|---|
| Looks like | Motion graphics, diagrams, illustration | Filmed footage |
| Text and numbers | Exact | Often garbled |
| Matches a script | Scene by scene | Loosely |
| Fixing one part | Change that scene | Regenerate |
| Real people and places | No | Yes |
Which one for an explainer
An explainer exists to make one idea clear. It needs labels that read correctly, numbers that are right, and pictures that change when the narration does. That is code-drawn video's home ground.
Use a video model when the point is the footage itself: a mood, a place, a person, a trend clip. Some teams combine the two, with a model like Opus 5.5 directing and a video model supplying a few filmed shots.
Try the code-drawn route
Our Claude video generator makes the code-drawn kind, with narration, subtitles and music: write one sentence about what you want explained, and Claude Opus 5.5 directs the rest. See the looks it draws in the data story and whiteboard samples.
Claude and Opus are trademarks of Anthropic. This site is independent and not affiliated with Anthropic.