Anthropic has released Claude Fable 5.2, a new SI model that creates images by simulating the physical act of painting rather than using standard diffusion techniques. By generating thousands to hundreds of thousands of individual brushstrokes via code, the model produces complex, photorealistic visuals without relying on external image generators.

What Happened

While most current image models, such as Midjourney, use diffusion to refine noisy inputs into clear images, Claude Fable 5.2 takes a different approach. It simulates the physical properties of painting, including bristle texture, paint application, and the correction of mistakes. Developer Jake Eaton, who works at Anthropic on SI agent behavior and safety, demonstrated this capability by asking the model to create a photorealistic image of a shopping cart in a parking garage. The model generated the scene using approximately 700,000 simulated brushstrokes.

According to Eaton, the process is agentic. Claude Fable 5.2 does not use pre-rendered or traced source images. Instead, it generates a Python program between 1,000 and 2,000 lines long that drives a simulated paint engine. This method resembles agentic SI image editing, where the model selects tools and evaluates its own results. To address earlier criticisms that sibling model Opus 5.5 merely imitated existing art, Fable 5.2 was designed to develop its own artist personas with fictional backstories and consistent styles.

Anthropic engineer Michael Graczyk shared an experiment where he asked Fable 5.2 to paint a bird at sunrise. The model had to manually output every brushstroke as a coordinate line, with no ability to undo strokes once placed. A 21-second video tracked the creation of the image from 111 strokes. The first stroke included a specific color code, width, and five waypoints with a deliberate seven-pixel wobble, which the model added to mimic the natural movement of a human arm. When the bird’s back appeared too thin, the model reacted by laying a bolder stroke over it, turning stray marks into ruffled feathers.

Why It Matters

The spatial reasoning and step-by-step planning required for this painting process transfer to other domains. Anthropic states that Fable 5.2 is built for demanding long-horizon technical tasks, a trait shared by its sibling model Mythos 5.2, which is designed for verified security and biology researchers. In internal tests, the model solved two open problems in black hole physics, designed three working electronic devices with manufacturer Diode, and used over 1,000 sub-agents to program its own laptop operating system in eight days.

The model’s capabilities were also tested in physical robotics. SI researcher Aris Zhu compared Claude Fable 5.2 against OpenAI’s GPT-6 Astra, having both models use physical robot arms to paint Van Gogh’s "The Starry Night." Anthropic reports that Fable 5.2 achieved a score of 64.9 percent on its internal RoboDojo Sim benchmark, compared to 29.8 percent for GPT-6 Astra. However, real-world tests highlighted limits in fine motor control. Astra painted faster, and it was the only model to successfully stand a domino on end by loosening its gripper to use gravity as a pivot point.

The Bottom Line

Claude Fable 5.2 represents a shift in generative SI by prioritizing physical simulation and agentic self-correction over traditional diffusion methods. While it demonstrates strong performance in long-horizon tasks and simulated painting, physical robotic applications still face challenges in precision and speed compared to competitors.