Anthropic released Claude Fable5.2 and Mythos5.2 on November 9 local time. The company has no traditional text-to-image model, yet Fable5.2 can use code to simulate a brush painting stroke by stroke on a virtual canvas, recreating brush texture, paint wetness, and canvas feel, and it can correct mistakes. Unlike diffusion models such as Midjourney, which generate images from noise, it breaks painting down into brushstroke commands and executes them itself. Each painting is a Python program, usually 1,000 to 2,000 lines long, and it does not rely on an image model or a pre-rendered base image.
Employee Jake Eaton demonstrated on X that after asking for a photorealistic piece, the model painted a shopping cart in a parking lot, using about 700,000 brushstrokes. In response to criticism that Opus5.5 mimicked existing styles, he had the model invent a fictional artist persona and adjust the brush based on text feedback.
