Anthropic's Claude Fable 5.2 paints images using Python code
Instead of using standard diffusion, the new model simulates physical brushstrokes to create art and solve complex technical problems.

The short version
- Fable 5.2 generates images by placing thousands of individual brushstrokes via code rather than using diffusion models.
- The model simulates physical properties like bristle texture, paint application, and the inability to undo a stroke.
- It scored 64.9 percent on Anthropic's internal RoboDojo Sim benchmark, beating OpenAI's GPT-6 Astra's 29.8 percent.
- The model has been used to solve black hole physics problems and design electronic devices with circuit board manufacturers.
Why it matters
This shift from 'generating' pixels to 'simulating' actions allows AI to reason about physical constraints and mistakes. It suggests a path toward robots that can plan complex, multi-step tasks by anticipating how physical materials will behave.
Simulating the artist's hand
Anthropic has introduced Claude Fable 5.2, a language model that creates images by simulating the physical act of painting. Unlike popular tools like Midjourney, which use diffusion to refine noisy images into clear pictures, Fable 5.2 writes Python code to place individual brushstrokes on a virtual canvas.
The model accounts for the physics of the medium, including the texture of bristles and how wet paint interacts with a surface. Because there is no undo button in this simulation, the AI must commit to every stroke it makes. This requires the model to plan its actions carefully to achieve a desired result.
In one demonstration, the model created a photorealistic shopping cart in a parking garage using roughly 700,000 simulated strokes. It also produced a bird painting using 111 coordinate sequences. To ensure consistency, the model can develop its own artist personas with specific backstories and styles.
Correcting mistakes in real-time
A key feature of Fable 5.2 is its ability to observe its own work and react to errors as a human painter would. When the model identifies a mistake, such as a stroke that is too thin or a smear from a paint tube, it generates new code to paint over the defect.
For example, when painting a bird, the model noticed the back looked too scraggly and laid a bolder stroke over it. The resulting marks from the original mistake were transformed into what looked like ruffled feathers. This ability to self-correct demonstrates a high level of spatial reasoning.
The model also handles complex spatial requests, such as having two figures paint with their outer hands so their inner shoulders can touch. By looking at the canvas between passes, the AI evaluates its progress and adjusts its strategy to meet the specific requirements of the prompt.
From canvas to complex physics
The reasoning capabilities of Fable 5.2 extend far beyond creating art on a virtual canvas. Anthropic designed the model for demanding long-horizon technical tasks that require step-by-step planning. It has already been used to solve two open problems in black hole physics and to design three working electronic devices.
The model's ability to manage complex workflows was further demonstrated when it used more than 1,000 sub-agents to program its own laptop operating system in eight days. These tasks require the model to understand sequences of actions and the consequences of each step within a technical framework.
This underlying logic is what allows the model to handle the
Overcoming imitation with agency
The development of Fable 5.2 follows earlier efforts with the sibling model Opus 5.5. Those previous attempts faced criticism because the model appeared to be merely imitating existing art rather than creating something original. To address this, Anthropic engineers moved toward a more agentic approach where the model develops its own unique artist personas.
By providing these personas with fictional backstories and consistent styles, the model can produce distinct work. This shift ensures that the AI is not just copying a style but is acting as an agent with a specific creative identity. This distinction is important for moving beyond simple imitation toward more complex, goal-oriented behavior.
Distinguishing from video generation
Fable 5.2 differs significantly from current methods used for AI-generated videos. Models like Flux 3 create video content by generating image sequences and maintaining temporal consistency directly. In contrast, Fable 5.2 builds its results through a sequence of manual actions, where each stroke is a commitment that cannot be undone once it is placed on the canvas.
This difference highlights the model's focus on physical simulation over direct visual output. While video models prioritize how a scene changes over time, Fable 5.2 prioritizes the logic of how a physical object is constructed. This makes it a unique tool for tasks requiring a deep understanding of how physical materials behave and interact with one another.
Real-world robotics and physics
While Fable 5.2 excels in simulated environments, it faces hurdles in real-world robotic control. In tests involving a physical robot arm painting a famous masterpiece, the model struggled with fine motor control compared to its competitor. For instance, OpenAI's model was able to stand a domino on end by slowly loosening its gripper to use gravity as a pivot point, a level of physical nuance the current version has yet to master.
The model's performance on the RoboDojo Sim benchmark shows its strength in planning, but the transition to physical hardware remains a challenge. Researchers are still working to bridge the gap between the model's high-level spatial reasoning and the precise, delicate movements required for real-world interaction. This ongoing work will determine how well these models can eventually operate in physical spaces.
What to watch
- How Fable 5.2's spatial reasoning improves in real-world robotic motor control.
- Whether the 'painting by code' method becomes a standard for other types of physical simulations.
Try it
Look for examples of 'agentic' AI workflows where a model uses tools to complete a task rather than just generating a final text response.
Sources and links
- OfficialFable 5.2 anthropic.com
- PostJake Eaton x.com
- PostJake Eaton / X x.com
- ReportingThe Decoder: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by The Decoder · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
