Claude Fable 5.2 approaches image making as a sequence of painting decisions. Instead of generating a finished picture in one pass, it writes code that directs simulated brushstrokes across a virtual canvas, then examines the result and paints over some mistakes.
A painting assembled through code
Many image models use diffusion, gradually turning noise into a clearer image. Fable 5.2 takes a different route: it simulates aspects of painting, including bristle texture, the application of paint and the need to cover unwanted marks.
Anthropic developer Jake Eaton demonstrated the method on X. For a request to make a painting difficult to distinguish from a photograph, Fable 5.2 depicted a shopping cart in a parking garage using roughly 700,000 simulated brushstrokes.
According to Eaton, the system does not rely on an external image model, a pre-rendered picture or a traced source image. Each painting is driven by a Python program of 1,000 to 2,000 lines that operates a simulated paint engine. That makes the result the product of a long series of choices about where and how to apply paint.
Each stroke becomes a decision
Anthropic engineer Michael Graczyk shared a chat log and video of a separate experiment: Fable 5.2 painting a bird at sunrise. The model had to represent the physics of bristles, wet paint and canvas texture in code, and describe every brushstroke as a line of coordinates.
There was no undo option. Once placed, a stroke remained on the canvas, so the model had to work with what it had already drawn. The resulting 21-second video follows the painting as it emerges from exactly 111 strokes.
For the opening stroke, Fable 5.2 specified the color code #30346a, a width of 90 pixels and five slightly wobbling waypoints. The model explained that the roughly seven-pixel wobble was intentional: a human arm does not draw a perfectly straight line either.
It also reviewed the canvas between passes. When the bird's back looked too thin and scraggly, the model added a stronger stroke over it. Marks at the edges then read as ruffled feathers. In another painting, showing two people at dusk, a lighting pass smeared paint tubes and a cup. Claude repainted those objects afterward, and arranged for both figures to paint with their outer hands so their inner shoulders could touch as requested.
From imitation to a developing style
Earlier efforts by the sibling model Opus 5.5 drew criticism for imitating existing art. Eaton's response was to have Fable 5.2 develop artist personas with fictional backstories and consistent styles. He also steers the model with comments as it works.
The method calls for several thousand brushstrokes per painting, so the result depends on sustained planning as well as local corrections. In practical terms, the model must account for how one mark affects the image and decide whether a later stroke can repair an earlier choice.
Virtual skill meets physical limits
The same step-by-step planning may matter beyond a simulated canvas. Anthropic describes Fable 5.2 as built for demanding long-horizon technical tasks. The source also says that, in internal tests, it solved two open problems in black hole physics, designed three working electronic devices with circuit board manufacturer Diode, and used more than 1,000 sub-agents to program its own laptop operating system in eight days.
AI researcher Aris Zhu compared Fable 5.2 with OpenAI's GPT-6 Astra using physical robot arms to paint Van Gogh's "The Starry Night." Anthropic reports Fable 5.2 scored 64.9 percent on its internal RoboDojo Sim benchmark, compared with 29.8 percent for GPT-6 Astra.
The real-world comparison points to a gap between benchmark results and dexterity. Astra painted faster, and so far only OpenAI's model has managed to stand a domino on end, slowly loosening its gripper so gravity could act as a pivot. The painting experiments show how code can turn image creation into a visible process; the robot test shows that translating planning into precise physical movement remains a separate challenge.