THE CRUNCH

Anthropic's Claude Fable 5.2 can produce complex, even photorealistic paintings without using an image generator at all. Instead of diffusion, the technique behind models like Midjourney, it writes Python programs of roughly 1,000 to 2,000 lines that drive a simulated paint engine, placing hundreds of thousands of individual brushstrokes on a virtual canvas. The simulation covers bristle texture, how paint goes on, and mistakes the model then has to paint over.

Developer Jake Eaton, who works at Anthropic on language model values and safety research, showed what the approach can do. Asked to paint something as hard to distinguish from a photograph as possible, the model produced a shopping cart in a parking garage using roughly 700,000 simulated brushstrokes. According to Eaton, Claude works without an external image model and without pre-rendered or traced source images.

Anthropic engineer Michael Graczyk shared a separate experiment in which the model painted a bird at sunrise, outputting every brushstroke as its own coordinate line with no way to undo strokes once placed. The 21-second video tracks the painting from exactly 111 strokes, starting with a stroke in colour #30346a, 90 pixels wide, with five slightly wobbling waypoints. Fable 5.2 said the roughly seven-pixel wobble was deliberate, because a human arm doesn't draw a perfectly straight line either.

The model also reacts to its own errors between passes. When the bird's back came out too thin and scraggly, it laid a bolder stroke over the stray marks so they read as ruffled feathers. In a second, 280-stroke piece of two people painting at dusk, an accidental smear of paint tubes and a cup was repainted after the fact. Eaton has also had the model develop its own artist personas with fictional backstories and consistent styles, after earlier attempts with the sibling model Opus 5.5 drew criticism for merely imitating existing art.

Anthropic positions Fable 5.2 as built for demanding long-horizon technical tasks, and says internal tests saw it solve two open problems in black hole physics, design three working electronic devices with manufacturer Diode, and use more than 1,000 sub-agents to program its own laptop operating system in eight days. In a real-world comparison by AI researcher Aris Zhu, Fable 5.2 and OpenAI's GPT-6 Astra each painted Van Gogh's The Starry Night with physical robot arms: Anthropic reports 64.9 percent on its internal RoboDojo Sim benchmark against 29.8 percent for Astra, though Astra painted faster and is so far the only one of the two to stand a domino on end.