Anthropic released Claude Fable 5.2, a language model that paints rather than generates pixels, placing individual brushstrokes on a virtual canvas via Python code while simulating bristle texture and paint application. Demos include a parking garage scene built from roughly 700,000 strokes and a bird painting from 111 coordinate sequences, with deviations painted over afterward. It scored 64.9% on Anthropic's internal RoboDojo Sim benchmark versus 29.8% for OpenAI's GPT-6 Astra, but trailed in speed and fine motor control in a real robot-arm test.
Termsslug: robodojo-sim
RoboDojo Sim
Anthropic's internal simulation benchmark for evaluating robotic manipulation capability.
Curated Intelligence
The DecoderT2·Models
78 pts
机器之心 (微信公众号)T2·Models
78 pts
Anthropic released Claude Fable 5.2 and Mythos 5.2, with Fable 5.2 scoring 64.9% on the RoboDojo Sim robotics benchmark, up 37.3 points from the prior 27.6% and ahead of Claude Opus 5.5 and GPT-6 Astra as well as most robotics-specific systems. Official demos show real-robot manipulation tasks, and former Tesla Optimus researcher Bingjie Tang has joined the company, signaling Anthropic's push of general models into embodied AI.