Anthropic's Claude Fable 5.2 generates images by programmatically simulating physical painting, outputting Python code that directs a virtual paint engine to place individual brushstrokes with modeled bristle texture, wet paint dynamics, and canvas interaction. In a demonstration by Anthropic developer Jake Eaton, the model produced a photorealistic parking garage scene using approximately 700,000 strokes, operating without external image models or traced sources. Each painting corresponds to a 1,000–2,000 line Python program; the model cannot undo strokes, forcing it to correct errors by overpainting, as shown when it thickened a bird's back to resemble ruffled feathers. In a separate test logged by engineer Michael Graczyk, Fable 5.2 painted a bird at sunrise in 111 coordinate-specified strokes, deliberately adding a seven-pixel wobble to mimic human motor variance, and later repainted objects accidentally smeared by a lighting effect. The model also demonstrated spatial reasoning by positioning two figures to paint with outer hands so inner shoulders touched. Beyond virtual canvases, Anthropic reports Fable 5.2 scored 64.9% on the internal RoboDojo Sim benchmark versus 29.8% for OpenAI's GPT-6 Astra, and solved two open black hole physics problems, designed three working electronic devices with manufacturer Diode, and programmed a laptop OS using over 1,000 sub-agents in eight days. However, in a real-world robotic arm test painting "The Starry Night" conducted by researcher Aris Zhu, GPT-6 Astra executed faster and exhibited superior fine motor control, remaining the only model to successfully balance a domino on end by exploiting gravity via gradual gripper release.

Read original