Without invoking an image model, Claude Fable5.2 paints a shopping cart with 700,000 brushstrokes
Anthropic released Claude Fable5.2 and Mythos5.2 on November 9. Fable5.2 has no traditional text-to-image model, yet it can paint stroke by stroke with code on a virtual canvas, reproducing brush texture, paint wetness, and canvas texture, and can correct mistakes. Unlike diffusion models such as Midjourney that generate images from noise, it breaks painting down into brushstroke instructions for execution; each painting is a Python program of about 1,000 to 2,000 lines and does not rely on images