Claude Fable 5.2 paints images with code and 700,000 brushstrokes
In a robot test by Anthropic, Fable 5.2 got 64.9 percent and GPT-6 Astra got 29.8 percent.
Anthropic released the language model Claude Fable 5.2. It makes images with Python code that puts brushstrokes on a virtual canvas. Jake Eaton of Anthropic showed an image of a cart in a garage that a person can take for a photo. It has about 700,000 brushstrokes. Tools such as Midjourney use diffusion, which makes a clear image from noise, but Fable 5.2 copies the steps of a painter.
How it works
Each image is a Python sequence of 1,000 to 2,000 lines of code that controls a simulated paint engine. Eaton tells that the model uses no other image model and no traced source images. In a bird test, the model could not remove a brushstroke. Between steps it looked at the canvas and painted on its errors.
The numbers
Anthropic reports 64.9 percent for Fable 5.2 on the RoboDojo Sim benchmark, against 29.8 percent for GPT-6 Astra from OpenAI. Aris Zhu, a researcher, did a test of the models with a robot arm that paints The Starry Night. Astra had more speed and more control of small movements.
What is not known
The source gives no test of the images against diffusion tools. The robot score comes from a test by Anthropic. The examples of the images come from persons who work at Anthropic.
Insight: The best example and the best score come from Anthropic
Michael Graczyk and Eaton, who show the images, work at Anthropic. The 64.9 percent score is also from Anthropic. On 9 November, Anthropic gave no test scores for code or vision. This reads as a wish of Anthropic to show the model with examples. The one robot arm test, by Zhu, showed more speed for OpenAI. Test it on your tasks.
Sources
Posted