Anthropic’s Claude Fable 5.2 paints images by outputting Python brushstroke programs
Anthropic’s new model Claude Fable 5.2 produces images by emitting Python code that places individual simulated brushstrokes on a virtual canvas, modeling bristle texture, paint application and self-correction; examples ranged from a 700,000‑stroke parking garage to a 111‑stroke bird. In internal RoboDojo Sim tests Fable 5.2 scored 64.9% versus 29.8% for OpenAI’s GPT-6 Astra, though real‑world robot arm tests showed GPT-6 Astra was faster and better at some fine-motor tasks.