Claude Fable 5.2 paints images through Python code, using about 700,000 simulated strokes in one piece
Anthropic's Claude Fable 5.2 produces images by writing Python programs that drive a simulated paint engine, with no external image model. Anthropic reports 64.9 percent on its internal RoboDojo Sim benchmark against 29.8 percent for OpenAI's GPT-6 Astra, though a real-world robot arm test showed limits in speed and fine motor control.
Anthropic's Claude Fable 5.2 creates images by writing Python code that places individual brushstrokes on a virtual canvas, rather than using diffusion. In one demonstration, the model produced a photorealistic parking garage scene from roughly 700,000 simulated strokes. Anthropic reports that Fable 5.2 scored 64.9 percent on its internal RoboDojo Sim benchmark, against 29.8 percent for OpenAI's GPT-6 Astra.
How the painting works
Image models such as Midjourney typically use diffusion, which refines a noisy image into a clear one. Fable 5.2 instead simulates the physical painting process, including bristle texture, paint application and mistakes it then paints over.
Anthropic's Jake Eaton, who contributes to research on language model values, safety mechanisms and agent behavior, demonstrated the approach on X. According to Eaton, Claude works without an external image model and without pre-rendered or traced source images. Each piece is a Python program of 1,000 to 2,000 lines driving a simulated paint engine. Asked for a painting as hard to distinguish from a photograph as possible, the model produced a shopping cart in a parking garage using roughly 700,000 strokes.
Earlier attempts with the sibling model Opus 5.5 drew criticism for imitating existing art. Eaton responded by having Fable 5.2 develop its own artist personas with fictional backstories and consistent styles.
No undo button
Anthropic engineer Michael Graczyk shared a chat log and video of an experiment in which Fable 5.2 painted a bird at sunrise. The model could only use code to simulate bristles, wet paint and canvas texture, and had to output every stroke as its own coordinate line, with no way to undo placed strokes. The resulting 21-second video shows the painting taking shape from exactly 111 strokes.
For the first stroke, the model output the color code #30346a, a width of 90 pixels and five slightly wobbling waypoints. Fable 5.2 said it added a wobble of roughly seven pixels deliberately, because a human arm does not draw a perfectly straight line. When the bird's back came out too thin, the model laid a bolder stroke over it, so stray bristle marks now read as ruffled feathers.
In a second piece of 280 strokes, showing two people painting at dusk, a late lighting effect accidentally smeared paint tubes and a cup on the table. Claude repainted those objects. It also had both figures paint with their outer hands so their inner shoulders could touch, as the prompt requested.
Benchmark and robot test
According to Anthropic, Fable 5.2 is built for demanding long-horizon technical tasks, as is Mythos 5.2, a version designed for verified security and biology researchers. In internal tests, Anthropic says Fable 5.2:
- solved two open problems in black hole physics
- designed three working electronic devices with circuit board manufacturer Diode
- used more than 1,000 sub-agents to program its own laptop operating system in eight days
Anthropic reports 64.9 percent for Fable 5.2 on RoboDojo Sim, its internal benchmark, compared with 29.8 percent for GPT-6 Astra. These are company-reported figures from an internal test.
AI researcher Aris Zhu separately had Fable 5.2 and GPT-6 Astra each use a physical robot arm to paint Van Gogh's "The Starry Night." Astra painted faster, and so far only OpenAI's model has managed to stand a domino on end, by slowly loosening its gripper to use gravity as a pivot point.
Pricing, context window and parameter count for Fable 5.2 are not yet disclosed.
What this means
The demonstrations show a language model handling image creation as a sequential, committed planning task: it writes code, inspects the canvas between passes and corrects its own errors. Because the model cannot undo strokes, the repainting examples (the bird's back, the smeared table objects) are evidence of self-correction rather than regeneration.
The benchmark comparison should be read with caution. RoboDojo Sim is Anthropic's internal benchmark, and the figures are reported by the company. The independent robot arm test by Zhu points the other way on speed and fine motor control, so the simulated lead did not carry over fully to physical hardware.
Related Articles
Gemini 4 Carbon tops MathArena at 85.4%, ahead of Claude Fable 5.2 and GPT-6.1 Sol
Google's Gemini 4 Carbon (high) leads the MathArena math benchmark with an expected performance of 85.4%, ahead of Anthropic's Claude Fable 5.2 (max) at 82.8% and OpenAI's GPT-6.1 Sol (max) at 80.6%. With the reasoning settings used in testing, Carbon costs roughly 15 times as much as GPT-6.1 Sol.
SPU wet-lab benchmark: top model Claude Opus 5.5 scores 0.486 and beats chance on only half of tasks
UniPat AI has released Science Prediction Universe (SPU), a benchmark of 116 tasks and 4,207 wet-lab experiments. Claude Opus 5.5 posted the highest average score, 0.486, but beat the random-ranking threshold on only half of its tasks. Raising GPT-6 Astra's reasoning effort multiplied runtime and cost without a visible score gain.
Claude Fable 5.2 scores 65 on Artificial Analysis index but costs about 2x Opus 5.5 per task
Anthropic announced Claude Fable 5.2, its top model, on November 9 (US time). ITmedia tested it against Opus 5.5 and Sonnet 5.5 on two office tasks and found the best-structured output, at the highest cost.
Anthropic tells Claude Opus 5.5 users to drop "think hard" prompts and start at Medium effort
Anthropic says users should remove phrases like "think hard" from prompts for Claude Opus 5.5, which reasons adaptively on its own. The company also recommends starting at the default Medium effort level, which it says matches Opus 5 on High in internal tests.
Comments
Loading...