Technology

Claude Fable 5.2 Paints Stroke by Stroke With Code, Not an Image Model

Anthropic released Claude Fable 5.2, which does not call an image-generation model but writes Python code to simulate brushes and paint stroke by stroke; one shopping-cart piece used about 700,000 simulated strokes.

Published · Updated · Source: IT之家 · 机器人筛选 · AI translation
Illustration
IT Home, November 16: Anthropic released the Claude Fable 5.2 and Claude Mythos 5.2 models on November 9 local time. Fable 5.2 targets complex programming, scientific research and engineering tasks, and also demonstrates an "AI image generation" approach different from conventional text-to-image: instead of directly calling an image-generation model, it writes code to simulate brush movement on a virtual canvas and completes paintings stroke by stroke, simulating brush texture, paint wetness and canvas feel, and correcting errors with new strokes. A work shown by Anthropic employee Jake Eaton is a shopping cart in a parking lot, using about 700,000 simulated strokes; each painting is actually a 1,000- to 2,000-line Python program. An experiment by engineer Michael Graczyk required the model to simulate brush physics only through code, with strokes that cannot be undone once placed; a sunrise bird work used 111 strokes and was recorded in a 21-second video. Anthropic says the model can coordinate multiple sub-agents and repeatedly verify results; in internal tests it solved two open problems in black hole physics, worked with Diode to design three working electronic devices, and coordinated more than 1,000 sub-agents to write an operating system for a laptop in 8 days. Robotic-arm tests by researcher Aris Zhu showed Fable 5.2 scored 64.9% on the internal RoboDojo Sim benchmark versus 29.8% for GPT-6 Astra, though fine motor control in real robotic arms still has shortcomings. This information comes from the original source and was organized by AI.
Read the original source

Related news