Without calling an image model, Anthropic Claude Fable 5.2 uses Python code to simulate a real paintbrush painting
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on local time November 9. Among them, Fable 5.2 does not call an image generation model; instead, it writes 1,000 to 2,000 lines of Python code to simulate brushes, paint, and canvas, completing a painting stroke by stroke. A work shown by Anthropic employee Jake Eaton used about 700,000 simulated brushstrokes, while an experiment by engineer Michael Graczyk used 111 strokes to complete a bird at sunrise; once a brushstroke is laid down, it cannot be undone. In Anthropic's internal RoboDojo Sim dual-arm robot simulation benchmark, Fable 5.2 scored 64.9%, while OpenAI GPT-6 Astra scored 29.8%, but actual robotic arm tests showed that Fable 5.2 still has shortcomings in fine motion control.
Why read this
AI score 70
The original article uses multiple sets of experimental details to explain how the model paints stroke by stroke in code and self-corrects, and this planning ability can be observed in relation to robot control.
Topics
Body · AI translation
Add new crowd-testing promotional copy Add new crowd-testing promotional copy end Add App-exclusive traffic guidance
IT Home, November 16 news: Anthropic released its latest Claude Fable 5.2 and Claude Mythos 5.2 models on local time November 9.
Among them, Fable 5.2 not only targets complex programming, scientific research and engineering tasks, but also demonstrates an unusual kind of "AI image generation" capability (IT Home note: Anthropic still has no traditional text-to-image model; the commonly seen DeepSeek and Claude memes are all generated by other models).
Unlike the painting approach of traditional image generation models: Fable 5.2 can simulate brushes, paint and canvas by writing code, completing a painting stroke by stroke, rather than directly calling an image generation model to output a finished product.
Fable 5.2 does not directly generate images; instead, by writing code, it simulates the movement of real brushes on a virtual canvas, completing a painting stroke by stroke, and can even simulate brush texture, paint wetness and canvas texture, and correct errors with new strokes after discovering them.
Currently, mainstream AI image generation tools such as Midjourney usually use diffusion models, generating images by gradually processing random noise. Fable 5.2 adopts another approach, breaking the painting process down into a large number of independent stroke instructions, which the model plans and executes itself.
Anthropic employee Jake Eaton demonstrated this capability on X. He asked Fable 5.2 to create a work that was as photorealistic as possible, and Claude ultimately painted a shopping cart in a parking lot, using about 700,000 simulated strokes.
According to Eaton, these works do not rely on external image models, nor do they use pre-rendered or traced base images. Each painting is actually a Python program written by Claude, with code usually 1,000 to 2,000 lines long, controlling different types of brushes through a simulated painting engine to complete the creation.
This approach also allows the model to develop a relatively independent artistic style. Previously, some paintings by Claude Opus 5.5 were criticized for imitating existing artistic styles too closely.
In response, Eaton began letting Fable 5.2 design its own artist personas with fictional backgrounds and stable creative styles, then guiding the model through text feedback based on the work's effect to adjust the brushes and painting engine.
Anthropic engineer Michael Graczyk also shared another set of experiments. He asked Fable 5.2 to paint a bird at sunrise, and restricted the model to simulating brush physical properties, wet paint and canvas texture only through code. This meant the model had to output each stroke individually as coordinate instructions, and once a stroke landed on the canvas it could not be undone.
Ultimately, a 21-second video recorded the model's process of completing this work from a blank canvas, using a total of 111 strokes. The first stroke used color #30346a, with a brush width of 90 pixels, and contained 5 slightly wavering path points. The model stated that this roughly 7-pixel wobble was deliberately designed, because human arms cannot draw absolutely straight lines either.
During multiple rounds of painting, the model observes the canvas and checks its own errors. For example, when the bird's back was painted too thin and the lines looked messy, it used thicker strokes to cover the original area, making the residual bristle marks at the edges look more like fluffy feathers.
In another dusk scene completed with 280 strokes, the model painted a scene of two figures painting together. Because adding lighting effects later accidentally blurred the paint tubes and cup on the table, Claude redrew these objects with new strokes. In addition, the model had the two figures each paint with the hand on the outer side of their body, so that their inner shoulders could touch each other as required by the prompt.
This ability to plan step by step, observe results and correct itself is also considered relevant to physical tasks such as robot control. Anthropic stated that Fable 5.2 targets complex, long-duration technical tasks, and can coordinate multiple sub-agents during execution and improve results through repeated verification.
According to internal test results published by Anthropic, Fable 5.2 once solved two open problems in the field of black hole physics, worked with circuit board manufacturer Diode to design three working electronic devices, and also coordinated more than 1,000 sub-agents to write an operating system running on a laptop within 8 days.
AI researcher Aris Zhu also tested Fable 5.2's actual robot control capabilities. He had Fable 5.2 and OpenAI's GPT-6 Astra separately control a physical robotic arm, attempting to paint Van Gogh's "The Starry Night".
In Anthropic's published internal RoboDojo Sim simulated robot benchmark test, Fable 5.2 scored 64.9%, while GPT-6 Astra scored 29.8%. Anthropic said the test evaluated the model's ability to control a simulated dual-arm robot.
However, the actual robotic arm test also showed that Fable 5.2 still has shortcomings in fine motion control. According to the experiment records shared by Zhu, GPT-6 Astra painted faster, and so far only OpenAI's model has successfully made a domino stand upright. Its method was to slowly release the robotic gripper, letting the domino use the gripper as a pivot, complete the rotation with the help of gravity and finally stand upright.article_adlist[
Related reading: ]article_adlist Statement: This is an exclusive article from Sina.com; unauthorized reproduction is prohibited. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin
Non-targeted 300*250 button end