AI Agent News/stories/anthropic-s-claude-fable-5-2-paints-images-stroke-by-stroke-in-python-code.md

Infrastructure

Anthropic's Claude Fable 5.2 paints images stroke by stroke in Python code

Anthropic's Claude Fable 5.2 creates images by placing simulated brushstrokes on a virtual canvas, with each painting produced by a Python program that drives a paint engine. Anthropic reports a higher score than OpenAI's GPT-6 Astra on an internal robotics benchmark, but a real-world robot arm test found Astra faster and better at fine motor control.

Source: the-decoder.com/anthropics-claude-fable-5-2-doesnt-generate-images-it-paints-them

Anthropic's Claude Fable 5.2 generates pictures by placing brushstrokes on a virtual canvas, according to developer Jake Eaton, who works at Anthropic. Each painting is a Python program of 1,000 to 2,000 lines that drives a simulated paint engine, and the model uses no external image model and no pre-rendered or traced source images. For a photorealistic parking garage scene, the model used roughly 700,000 simulated strokes.

Anthropic engineer Michael Graczyk shared an experiment in which the model painted a bird at sunrise without any undo function, outputting each stroke as its own coordinate line. The painting took 111 strokes. When the bird's back came out too thin, the model laid a bolder stroke over it, and in a second piece it repainted objects that a lighting effect had smeared.

Anthropic reports a score of 64.9 percent for Fable 5.2 on its internal RoboDojo Sim benchmark, compared with 29.8 percent for OpenAI's GPT-6 Astra. In a separate test by AI researcher Aris Zhu, both models used physical robot arms to paint Van Gogh's "The Starry Night", and Astra painted faster and was the only model to stand a domino on end.

Why it matters if you build agents

Developers building agents can see a model that plans step by step and corrects its own output across long tasks, with the feedback loop written entirely in code. Anthropic says the same spatial reasoning carries over to physical robots, which makes that claim worth testing in your own control pipelines. The real-world robot test showed weaker fine motor control than Astra, and the benchmark scores are Anthropic's own figures, so wait for independent results before relying on them.

Get stories like this every weekday

One short email with the day's agent tooling news.

Subscribe