正文 · AI 翻译
IT Home reported on November 16 that Anthropic released its latest Claude Fable 5.2 and Claude Mythos 5.2 models on November 9 local time.
Fable 5.2 is designed for complex programming, scientific research, and engineering tasks, and has also demonstrated an unusual “AI image generation” capability (IT Home notes that Anthropic still has no text-to-image model in the traditional sense; the familiar DeepSeek and Claude memes are generated by other models).
Unlike the way traditional image generation models paint, Fable 5.2 can write code to simulate brushes, paint, and canvas, completing a painting stroke by stroke instead of directly calling an image generation model to produce a finished image.
Fable 5.2 does not generate images directly. Instead, it writes code to simulate the movement of a real brush on a virtual canvas, completing paintings stroke by stroke. It can even simulate brush texture, paint wetness, and canvas texture, and correct mistakes with new strokes after detecting them.
Mainstream AI image generation tools such as Midjourney typically use diffusion models, generating images by progressively processing random noise. Fable 5.2 takes a different approach, breaking the painting process into a large number of separate brushstroke instructions that 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 brushstrokes.
According to Eaton, these works do not rely on an external image model and do not use a pre-rendered or traced base image. Each painting is actually a Python program written by Claude, typically 1,000 to 2,000 lines long, that controls different types of brushes through a simulated painting engine to create the work.
This method also allows the model to develop a relatively independent artistic style. Some paintings by Claude Opus 5.5 had previously been criticized for imitating existing artistic styles too closely.
In response, Eaton began having Fable 5.2 design artist personas with fictional backgrounds and consistent creative styles, then used written feedback on the resulting works to guide the model in adjusting its 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 physics, wet paint, and canvas texture through code. This meant the model had to output each stroke separately as coordinate instructions, and strokes could not be undone once applied to the canvas.
The final result was a 21-second video documenting the model completing the work from a blank canvas with a total of 111 strokes. The first stroke used the color code #30346a, a brush width of 90 pixels, and 5 slightly wavering path points. The model said the roughly 7-pixel wobble was intentional because a human arm cannot draw a perfectly straight line either.
Over multiple rounds of painting, the model observes the canvas and checks for its own mistakes. For example, when it painted the bird’s back too thinly and the lines looked messy, it used broader strokes to cover the original area, making the bristle marks left at the edges look more like fluffy feathers.
In another work, a dusk scene completed with 280 strokes, the model depicted 2 people painting together. When it accidentally blurred the paint tubes and cups on the table while adding lighting effects later, Claude repainted those objects with new strokes. The model also had each person paint with the hand on the outside of their body so their inner shoulders could touch, as required by the prompt.
This ability to plan step by step, observe results, and make corrections independently is also considered relevant to physical tasks such as robot control. Anthropic said Fable 5.2 is designed for complex technical tasks that run for extended periods, can coordinate multiple subagents during execution, and improves results through repeated verification.
According to internal test results released by Anthropic, Fable 5.2 has solved 2 open problems in black hole physics, worked with circuit board manufacturer Diode to design 3 functional electronic devices, and coordinated more than 1,000 subagents to write an operating system that runs on a laptop in 8 days.
AI researcher Aris Zhu also tested Fable 5.2’s ability to control a physical robot. He had Fable 5.2 and OpenAI’s GPT-6 Astra each control a physical robotic arm to try to paint Van Gogh’s The Starry Night.
In Anthropic’s internal RoboDojo Sim simulated robotics benchmark, Fable 5.2 scored 64.9%, while GPT-6 Astra scored 29.8%. Anthropic said the test assessed the models’ ability to control a simulated dual-arm robot.
However, the physical robotic arm tests also showed that Fable 5.2 still has shortcomings in fine motor control. According to the experimental records shared by Zhu, GPT-6 Astra painted faster, and so far only OpenAI’s model has successfully stood a domino upright. Its method was to slowly release the robotic gripper, using it as a pivot so gravity would rotate the domino until it stood upright.
Related reading:
“The world’s most powerful AI model: Claude Fable 5.2 arrives, surpasses GPT-6 Astra in programming and tops multiple benchmarks”
“An operating system built from scratch in 8 days: Anthropic’s most powerful Claude Fable / Mythos 5.2 AI models arrive, with API call costs down 20%”