Claude Fable 5.2 paints images with coded brushstrokes
Anthropic’s Claude Fable 5.2 creates images through a simulated painting process rather than a conventional diffusion system. The language model writes Python programs, typically 1,000 to 2,000 lines long, that control a virtual paint engine. The code models bristle texture, paint application and canvas behavior, while the model evaluates the result and adds corrective strokes.
In a demonstration by Anthropic developer Jake Eaton, the system produced a photorealistic shopping cart in a parking garage using about 700,000 simulated brushstrokes. In another experiment, it painted a bird at sunrise through 111 coordinate sequences. Each stroke had to be specified manually, and there was no undo function. The model adjusted later marks when proportions or details were wrong, including strengthening the bird’s back and repainting objects affected by an unintended lighting pass.
Anthropic reports that Fable 5.2 scored 64.9% on its internal RoboDojo Sim benchmark, versus 29.8% for OpenAI’s GPT-6 Astra. A physical-robot comparison showed limitations: Astra painted faster, and only Astra managed to balance a domino upright. Anthropic also says Fable 5.2 solved two open black-hole physics problems, designed three working electronic devices and used more than 1,000 sub-agents to program an operating system.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.