Anthropic's Claude Fable 5.2 doesn't generate images, it paints them

Anthropic's new language model Claude Fable 5.2 produces images in an unusual way: instead of generating pixels, it writes Python programs that drive a simulated paint engine, placing individual brushstrokes on a virtual canvas with modeled bristle texture and paint application. Anthropic developer Jake Eaton demonstrated the capability on X, and engineer Michael Graczyk shared a chat log and video of a related experiment. The approach departs from the diffusion methods used by image models such as Midjourney, which refine a noisy image into a clear result over successive steps.
In Eaton's demonstration, Fable 5.2 was asked to make a painting as hard to distinguish from a photograph as possible. It produced a shopping cart in a parking garage using roughly 700,000 simulated brushstrokes. According to Eaton, the model works without an external image model and without pre-rendered or traced source images. Each finished piece is a Python program of 1,000 to 2,000 lines that drives the paint engine. Eaton described the process as resembling agentic image editing, in which a language model selects tools and evaluates the results.
An earlier attempt with the sibling model Opus 5.5 drew criticism for merely imitating existing art, Eaton said. He addressed that by having Fable 5.2 develop its own artist personas, complete with fictional backstories and consistent styles. The model requires several thousand brushstrokes per painting, and Eaton steers it with comments.
Graczyk's experiment asked the model to paint a bird at sunrise using only code that simulates the physics of bristles, wet paint and canvas texture. Fable 5.2 had to output every stroke manually as its own line of coordinates, with no way to undo a stroke once placed. The resulting 21-second video uses a counter to show the painting taking shape from exactly 111 strokes on a blank canvas. The first stroke specified the color code #30346a, a width of 90 pixels and five slightly wobbling waypoints. The model said the roughly seven-pixel wobble was deliberate, because a human arm does not draw a perfectly straight line either.
Between passes, the model inspected the canvas and reacted to its own errors. When the bird's back came out too thin and scraggly, it laid a bolder stroke over the area, so stray bristle marks along the edges read as ruffled feathers. In a second piece of 280 strokes showing two people painting at dusk, a late lighting effect accidentally smeared paint tubes and a cup on the table, and Claude repainted those objects afterward. The model also positioned both figures to paint with their outer hands so their inner shoulders could touch, as the prompt requested.
This is a meaningfully different bet from diffusion. A standard image model spends its compute on a pixel grid and has no notion of how a mark was made; here the language model spends its tokens on a literal sequence of actions, each one constrained by the stroke that came before. That constraint is the point. A plan that cannot be undone forces the model to reason about consequences, which is closer to the kind of long-horizon control Anthropic appears to be targeting than to image synthesis as such.
The trade-off is cost and speed. Hundreds of thousands of strokes, each emitted as code, is a lot of generated output for a single picture, and the fidelity depends on the paint engine as much as on the model. It is also unclear how this scales to the kind of volume a production image pipeline expects.
Anthropic frames Fable 5.2 as built for extremely demanding long-horizon technical tasks, the same positioning it gives Mythos 5.2, a version aimed at verified security and biology researchers. In internal tests, the company says, Fable 5.2 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 tested the model with real robotic arms, pitting Fable 5.2 against OpenAI's GPT-6 Astra; each attempted to paint Van Gogh's "The Starry Night." Anthropic reports a score of 64.9 percent for Fable 5.2 on its internal RoboDojo Sim benchmark, against 29.8 percent for GPT-6 Astra. The real-world comparison still shows limits in fine motor control: Astra painted faster, and only OpenAI's model has so far managed to stand a domino on end, which it did by slowly loosening its gripper to use gravity as a pivot point.
A benchmark gap of more than two to one in simulation, paired with a real-world speed disadvantage and a task the competitor has completed and Fable 5.2 apparently has not, is worth reading carefully. Simulations tend to reward planning and error correction, which is where this model seems strongest, while physical hardware rewards calibration and grip control, which depend on how the robot is driven. The gap likely reflects both model capability and how each system is integrated with the arm.
Why it matters: Developers building agents that must act over long sequences, whether in design tools, robotics or code, may find the more interesting signal here not in the paintings but in the step-by-step, self-correcting planning they demonstrate. My inference is that the technique is unlikely to replace pixel-based image generation on cost or throughput soon, but it offers a concrete testbed for measuring whether a model can commit to irreversible actions and recover from its own mistakes. Product teams evaluating agentic workflows should watch the RoboDojo results, while keeping in mind that real hardware has so far favored the competing model on speed and dexterity.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Anthropic's Claude Fable 5.2 doesn't generate images the way other AI models do. Instead of computing pixels, the language model writes Python code that places hundreds of thousands of simulated brushstrokes on a virtual canvas, complete with bristle texture and paint application. A single photorealistic parking garage scene took roughly 700,000 strokes, according to Anthropic employee Jake Eaton. The article Anthropic's Claude Fable 5.2 doesn't generate images, it paints them appeared first on The Decoder .