Models

Anthropic's Claude Fable 5.2 Paints With 700,000 Simulated Brushstrokes

Anthropic's Claude Fable 5.2 paints images stroke by stroke in Python — 700,000 for one photorealistic scene — and beat GPT-6 Astra on an internal robotics benchmark, 64.9 to 29.8 percent.

Anthropic's Claude Fable 5.2 doesn't generate images, it paints them
Anthropic's Claude Fable 5.2 doesn't generate images, it paints themKwami Fattah Al Sissi / unsplash
By James Calloway6 min read

Why it matters

  • Claude Fable 5.2 rendered a photorealistic parking garage scene from roughly 700,000 simulated brushstrokes.
  • Each painting is a Python program of 1,000 to 2,000 lines driving a simulated paint engine, with no undo option.
  • Fable 5.2 scored 64.9 percent on Anthropic's internal RoboDojo Sim benchmark versus 29.8 percent for OpenAI's GPT-6 Astra.
  • Anthropic says Fable 5.2 used over 1,000 sub-agents to program its own laptop operating system in eight days.
  • GPT-6 Astra painted faster in Zhu's robot arm test and remains the only model to stand a domino on end.

Anthropic's new language model Claude Fable 5.2 produced a photorealistic image of a shopping cart in a parking garage using roughly 700,000 simulated brushstrokes — placing each one individually through code, with no diffusion process and no image generator underneath. The demonstrations, published by Anthropic staff this week, show a fundamentally different approach to image creation than what powers Midjourney and most current models.

The distinction matters commercially and technically. Current image models rely on diffusion, a method that gradually refines a noisy starting image into a clear result. Fable 5.2 instead simulates the physical act of painting: bristle texture, how paint adheres, and the mistakes a painter must then paint over. If the results hold up, Anthropic is claiming a route to image generation grounded in physical reasoning rather than statistical denoising — a claim that, so far, rests almost entirely on curated internal demos.

How does the model actually paint?

Each artwork is a Python program of 1,000 to 2,000 lines driving a simulated paint engine, according to Jake Eaton, an Anthropic developer who works on language model values, safety mechanisms, and AI agent behavior. Eaton demonstrated the approach on X. He states that Claude works without an external image model and without pre-rendered or traced source images — meaning the model is not copying an existing picture and reproducing its strokes.

The method resembles agentic image editing, in which language models select tools and evaluate their own results in a loop. Eaton prompted Fable 5.2 to create a painting as hard to distinguish from a photograph as possible. The model chose a parking garage scene and rendered it stroke by stroke. That is the strongest version of Anthropic's claim; independent reproduction has not yet been published.

Earlier attempts with the sibling model Opus 5.5 drew criticism for merely imitating existing art. Eaton's response was procedural, not defensive: he had Fable 5.2 develop its own artist personas, complete with fictional backstories and consistent styles. The model needs several thousand brushstrokes per painting, and Eaton steers it mid-process with comments. Whether invented personas fully answer the imitation critique is a judgment the demonstrations leave to the viewer.

What happens when a stroke goes wrong?

Anthropic engineer Michael Graczyk shared a chat log and video of an experiment in which he asked Fable 5.2 to paint a bird at sunrise. The constraints were strict. The model could only use code to simulate the physics of bristles, wet paint, and canvas texture. It had to output every brushstroke manually as its own coordinate line, with no way to undo a stroke once placed.

That last constraint separates this method from AI-generated video, where models like Flux 3 produce image sequences and their temporal consistency directly. Without an undo button, Fable 5.2 must commit to every stroke.

The resulting 21-second video tracks the painting taking shape from exactly 111 strokes on a blank canvas, counted on screen. 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 the roughly seven-pixel wobble deliberately, because a human arm doesn't draw a perfectly straight line either.

Between passes, the model examined the canvas and reacted to its own mistakes the way a painter would. When the bird's back came out too thin and scraggly, the model laid a bolder stroke over it. The stray bristle marks along the edges now read as ruffled feathers — an error converted into a detail. That is a compelling demo of self-correction, though it is also a single, Anthropic-selected example.

A second piece makes the spatial reasoning harder to dismiss. In a 280-stroke painting of two people painting at dusk, a late lighting effect accidentally smeared the paint tubes and a cup on the table. Claude repainted those objects after the fact. The model also had both figures paint with their outer hands so their inner shoulders could touch, exactly as the prompt requested — a compositional constraint that requires modeling bodies in space, not just style.

Does this transfer beyond the canvas?

Anthropic says yes, and frames Fable 5.2's painting as a symptom of something larger. According to the company, the model is built for extremely demanding long-horizon technical tasks, as is Mythos 5.2, a variant designed for verified security and biology researchers. In internal tests, Anthropic reports that 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.

These figures come from Anthropic's own testing, with no published methodology or external verification attached to the announcements. The black hole and operating system claims in particular would benefit from documentation the company has not yet provided.

How does it stack up against GPT-6 Astra?

The one head-to-head number comes from Anthropic's internal RoboDojo Sim benchmark: Fable 5.2 scored 64.9 percent, against 29.8 percent for OpenAI's GPT-6 Astra. On its own benchmark, Anthropic's model more than doubles the competitor's score — the usual caveat about self-graded homework applies.

The real-world test was run by an outside researcher, AI researcher Aris Zhu, and it cuts both ways. Zhu pitted Fable 5.2 against GPT-6 Astra, with each model using a physical robot arm to paint Van Gogh's "The Starry Night." Fable 5.2 won on painting quality by Anthropic's account. But Zhu's comparison exposed limits in fine motor control: Astra painted faster, and so far only OpenAI's model has managed to stand a domino on end — doing so by slowly loosening its gripper to use gravity as a pivot point.

Speed and dexterity remain OpenAI's edge; planning and self-correction under physical constraints appear to be Anthropic's. Neither company has published enough detail to settle the trade-off independently.

The stakes extend past novelty painting. A model that plans thousands of sequential actions, monitors its own output, and repairs errors without an undo function is demonstrating the core competencies robotics, automated engineering, and long-horizon AI agents require. Anthropic's claims — two physics results, three working devices, an operating system in eight days — are calibrated to make exactly that argument. Until independent evaluations appear, the strongest verifiable fact remains that a language model painted a bird in 111 committed strokes and corrected its own mistakes along the way.

Original: x.com

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

604 articles

Related articles

  1. Adobe's EditAgent Makes AI Judge Its Own Photo Edits With Hard Numbers
  2. Elorian launches Glimpse, an AI model that looks again
  3. Google grounds Project Genie worlds in Street View imagery
  4. Google's Gemini 4 Argon scores 87% on embodied-reasoning benchmark
  5. Epic Games goes "AI first" on Disney project, Bloomberg reports

« Previous article