Claude Fable 5.2 just created an entire 4K CGI short using Python, Blender, RunPod, and ElevenLabs, starting from a single prompt.
The reported GPU rendering cost? Just $5.75.
In an experiment by Riley Brown, Claude was asked to create an animated room with a MacBook, some vibe-coding magic, and Clawd, Anthropic's little mascot.
Instead of generating video frames with a diffusion model, Claude wrote Python scripts to construct and animate an actual 3D scene inside Blender.
And the level of detail is remarkable.
Here's what Claude reportedly built:
• Procedural 3D environment: an entire room, desk, lighting setup, MacBook, and animated Clawd character, constructed through Python.
• Detailed MacBook: 77 individually modeled and animated keyboard keys, synchronized with the typing displayed on screen.
• A working screen animation: Claude built a web page resembling the Claude interface, captured it as 1,345 sequential images, and used them as animated textures on the laptop display.
• Frame-accurate keyboard animation: individual keys move when their corresponding letters appear on screen.
• 13 cinematic shots: camera compositions and movements generated programmatically.
• Animated lighting: a simulated sun moves from golden hour through nighttime and into sunrise.
• Cloud GPU rendering: Claude orchestrated a 4K render across 10 rented GPUs on RunPod.
• Rendering cost: approximately $5.75 for GPU rendering, completed in about 52 minutes.
• AI-generated audio: music and sound effects created using ElevenLabs.
• Automated post-production: color grading, titles, and final video assembly.
The interesting part isn't simply that Claude generated a good-looking animation.
It's that the entire production was built through code.
The geometry exists in Blender. The cameras are editable. The lighting can be modified. The animation can be adjusted and rendered again.
This is fundamentally different from generating a video directly with a text-to-video model.
For independent filmmakers, motion designers, and VFX artists, AI agents capable of controlling traditional 3D software could dramatically change production workflows.
Of course, $5.75 is the reported GPU rendering cost, not the total cost of model inference, development time, or all production services.
And one initial prompt doesn't necessarily mean the same results can be reproduced without existing context or further iteration.
Still, the technical possibilities are fascinating.
Original experiment by Riley Brown:
https://x.com/rileybrown/status/2119902109709537708
Blender:
https://www.blender.org/
RunPod:
https://www.runpod.io/
ElevenLabs:
https://elevenlabs.io/
Have you tried controlling Blender with Claude Fable 5.2?
Would you rather generate AI video directly or have an AI coding agent build a fully editable CGI scene?
#ClaudeFable52 #Blender3D #AIFilmmaking
YouTube / ABVX
