Two of my eight Shorts cost five times more than the rest, and both times it was my fault.
Some context first. I make animated YouTube Shorts with Claude Code and no video model. No Veo, no Sora, nothing billed per second. Claude writes the whole film as HTML and JavaScript, HyperFrames renders the frames in a headless browser, and ffmpeg stitches them into an MP4 on my own machine.
I log what every session would cost at API list prices. Here's what that log says.
The numbers
Between Oct 15 and Oct 23 I made 8 Shorts for my channel `reelstyleshd, each around 31 to 38 seconds.
Six of them came in between $4.10 and $6.90 each at Opus list prices. That includes the automatic check-and-fix rounds where Claude looks at its own frames and fixes what's off.
The other two cost $23 and $24. Those are the ones I threw away and rebuilt from zero.
For comparison, 35 seconds of Veo 3.1 Fast was listed at $3.50 to $4.20 when I checked in October. So on raw API price for a single clip, it's roughly a tie. I'm not going to pretend code is magically cheaper per clip.
(I actually run this on a flat Claude subscription, so I don't get a bill. The numbers above are what the API would charge.)
Where it actually wins
The film is code. So when a label is wrong or a character is in the wrong spot, I change a line and render again locally. Free. With a video model you pay again for the whole take.
That's also why the two expensive films hurt. I skipped the cheap checks and paid for a full rebuild instead of a small fix.
What I do now:
- Claude writes a short shot list first and I approve it before any code exists
- I ask for stills (
npx hyperframes snapshot) and look at them before the full render - I build on Sonnet and only switch to Opus for scenes with heavy drawing
A wrong frame in a still takes seconds to spot. Spotting it after the render costs you the render.
How long a render takes
From an empty folder to a finished MP4, on Sonnet, on a shared 6-core server: a 10 second Short took 3.8 minutes (median of 29 runs). A 30 second film took 14 minutes (median of 67 runs). No GPU involved.
The part nobody warns you about
Claude Code plus HyperFrames works without anything extra. But left alone, the films all look the same. Centered text on a gradient, shapes sliding in.
So I started writing style kits. A kit is a folder of plain files: palette, how characters are drawn, scene templates, motion rules, and a checklist Claude grades its own stills against. Claude reads the kit before it draws a single scene, and episode 10 looks like episode 1.
Because characters are drawn by functions, shot 7 matches shot 1 too. That's the thing video models still struggle with.
To be fair, this only works for drawn looks. Cartoons, cut paper, diagrams, motion graphics, toy-like 3D. If you need realistic people or places, use a video model.
Try it
I put 6 kits on GitHub under MIT: dark keynote, whiteboard explainer, doodle film, kinetic type, a stick-figure metaphor Short, and a countdown guessing quiz. They're the same kits buyers get, not cut-down versions.
In Claude Code:
shellreelstyles-free
/plugin marketplace add reelstyles/reelstyles-free
/plugin install reelstyles-free
`
With Codex, Cursor or Gemini CLI, just clone the repo and point your agent at a kit. You need Node.js, Chrome and FFmpeg.
If you want more, Starter is 30 styles for $19 and Full is all 121 for $49 for the first 20 buyers, then $79.
Free kit: https://github.com/reelstyles/reelstyles-free
Everything else: https://reelstyles.com
If you make something with it, drop it in the comments. I want to see what breaks.
Top comments (0)