Anthropic's Claude Fable 5.2 doesn't generate images, it paints them
Reported by The Decoder and MLQ.ai2 sources ↓
Anthropic's Claude Fable 5.2 makes paintings by writing code that simulates brushstrokes on a virtual canvas, rather than using an external image model, The Decoder reports.
Key points
- Fable 5.2 simulates paint, bristles and canvas texture with code
- A shopping-cart image used roughly 700,000 simulated brushstrokes, according to Anthropic developer Jake Eaton
- The model can inspect the canvas between passes and paint over mistakes
- Anthropic reports a 64.9% score on its RoboDojo Sim benchmark, versus 29.8% for OpenAI's GPT-6 Astra
Sources
Details
The Decoder reports that Anthropic developer Jake Eaton demonstrated Fable 5.2 making a photorealistic-looking shopping cart in a parking garage. Eaton said the model used roughly 700,000 simulated brushstrokes, without an external image model or a pre-rendered or traced source image.
Each painting is produced by a Python program of about 1,000 to 2,000 lines that directs a simulated paint engine. The model can inspect the canvas between passes and respond to errors; in one example, it painted over a thin, scraggly-looking bird's back with a heavier stroke.
A separate demonstration showed the model building a painting of a bird at sunrise from 111 manually specified strokes, in a 21-second video.
Written from coverage as of
Images
- The Decoder
- The Decoder
- The Decoder
- The Decoder
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh