No Image Model Needed: Claude Fable 5.2 Drew a Shopping Cart with 700,000 Brushstrokes
1 day ago / Read about 0 minute
Author:小编   

Anthropic has launched Claude Fable 5.2 and Mythos 5.2. The company does not have a traditional text-to-image model; instead, Fable 5.2 simulates a paintbrush stroke by stroke on a virtual canvas through code, breaking painting down into brushstroke instructions for execution. Each painting corresponds to over a thousand lines of Python code, does not rely on image models or pre-rendered base images, can reproduce various painting textures, and can correct its own errors. Demonstrations show that it can complete paintings of varying complexity and possesses planning and self-correction capabilities. The company states that it is designed for complex, long-duration tasks and can coordinate a large number of sub-agents to accomplish highly difficult work. Internal tests show that it outperforms GPT-6 Astra on related benchmarks, but fine motion control on real machines still has shortcomings.