Skip to main content
Claude Fable 5.2 AI paints a hyperrealistic shopping cart using 700,000 simulated brushstrokes, showcasing advanced generativ

Editorial illustration for Claude Fable 5.2 paints photo-like shopping cart with 700,000 simulated strokes

Claude Fable 5.2 paints photo-like shopping cart with...

• 3 min read

Anthropic's latest language model skips the usual trick for making pictures. Claude Fable 5.2 doesn't generate an image in one noisy-to-clear pass the way Midjourney or Stable Diffusion do. Instead it writes Python code that places individual brushstrokes on a virtual canvas, one at a time, simulating things like bristle texture and how wet paint actually spreads and layers.

Jake Eaton, an Anthropic developer who works on language model values and AI agent behavior, posted examples on X showing what this produces. A bird painting came from 111 coordinate sequences. A more demanding test, a photorealistic shopping cart sitting in a parking garage, took the model roughly 700,000 simulated strokes to finish, with Fable 5.2 painting over its own deviations as it went, the way a human painter corrects a stray line.

Anthropic also ran the model through its internal RoboDojo Sim benchmark, pitting it against OpenAI's GPT-6 Astra. The results, and how Fable 5.2 held up once the task moved from a screen to a physical robot arm, say a lot about where this stroke-by-stroke approach actually stands.

Anthropic's new language model Claude Fable 5.2 can create complex and even photorealistic paintings by using code to place hundreds of thousands of individual brushstrokes on a virtual canvas.

Why this matters

A model that paints stroke by stroke, correcting its own mistakes mid-canvas, is a different animal from a diffusion model guessing pixels in one pass. For developers, the interesting part isn't the shopping cart photo, it's the 700,000-stroke process behind it: Claude Fable 5.2 is reasoning about bristle texture and paint layering through Python code, which means the output is inspectable and editable in ways a black-box image generator never was. That opens real doors for tooling built around iterative correction rather than one-shot prompts.

For researchers, the 111-coordinate bird painting is the more telling demo. It suggests the model can compress a visual idea into a short instruction set and still execute something convincing, which says more about planning and abstraction than about rendering fidelity. We'd push back on reading too much into "indistinguishable from a photograph" claims from Anthropic's own internal testing.

Self-graded benchmarks always flatter the grader. Still, a language model simulating physical paint application, rather than calling out to a separate image model, is worth watching closely as a signal of where multimodal reasoning is headed next.

LIVE17:05Claude Fable 5.2 paints photo-like shopping cart with 700,000 simulated strokes