Anthropic does not have a Vincentian diagram model, but Claude Fable 5.2 has been drawing recently.
The way it is drawn is completely different from diffusion models like Midjourney. The diffusion model "washes" the image out of the noise step by step; Fable 5.2 is a Python program written to simulate the fall of a pen on a virtual canvas one by one, simulating the brush texture, paint humidity and canvas texture. Each painting is a piece of code, usually one or two thousand lines, without the aid of an image model or a pre-rendered basemap.
Anthropic employee Jake Eaton asked him to draw a photorealistic painting of a shopping cart in a parking lot, using about 700,000 strokes.
Another example gives a better idea of what it is thinking. Engineer Michael Graczyk asked it to draw a bird at sunrise. The rule is that each stroke can only output coordinate instructions, and the stroke cannot be undone. It was painted in 111 strokes, starting with a 90 pixel brush using the color #30346a. Halfway through the painting, he would stop and look at the canvas. He would find that the back of the bird was thin, and then cover it up with a thick brush.
Some people have previously criticized Opus 5.5 for only imitating ready-made painting styles. Eaton's response was to let the model first program a fictional painter's persona for itself, and then adjust the brushwork based on text feedback.
This matter is associated in the direction of robotics within Anthropic: planning, execution, observation, correction, which is exactly the cycle that a robotic arm needs. Researcher Aris Zhu asked Fable 5.2 and GPT-6 Astra to control the robot arm to draw "Starry Night" respectively. The internal RoboDojo Sim benchmark was 64.9% versus 29.8%.
The situation is different on the real machine: fine movement control is not enough, GPT-6 Astra draws faster, and currently only OpenAI's model can stand up dominoes. I won in the simulation, but not on the real machine.