Anthropic’s Claude Fable 5.2 can make detailed pictures by writing code that places individual brushstrokes on a simulated canvas. Instead of using a conventional image generator, the language model produces Python programs that control paint, bristle texture and the position of each stroke.

Anthropic developer Jake Eaton demonstrated a parking-garage scene assembled from roughly 700,000 simulated strokes. He said the model used no external image model and no traced or pre-rendered source image. Each work is driven by a program of about 1,000 to 2,000 lines, and Eaton guides later attempts with written feedback.

The process differs from diffusion, the common technique in tools such as Midjourney that gradually turns noise into an image. In another experiment, Anthropic engineer Michael Graczyk asked the model to paint a bird at sunrise without an undo function, forcing it to work around earlier marks as a physical painter might. These are employee demonstrations rather than a benchmark or a broadly described product release, so they do not establish that the method is faster or more useful than existing image models. They do show that a language model can construct complex visual work through an inspectable program rather than direct pixel generation.