A picture, written out longhand
Diffusion image models start from noise and refine it until a picture appears. What Eaton describes is the opposite direction. Fable 5.2 writes a Python program, between 1,000 and 2,000 lines, that runs a simulated paint engine: bristle texture, how paint lands on canvas, and mistakes that then have to be painted over. The model writes the program; the program places the strokes. The picture is the side effect.
The headline example is a shopping cart in a parking garage, made with roughly 700,000 simulated brushstrokes after Eaton asked for something as hard to distinguish from a photograph as possible. Two qualifiers belong next to that number. First, it is Eaton's account of his own experiment, shared on X and reported by The Decoder; that account provides no independent replication or audit of the setup. Second, the result does not actually pass as a photograph up close: The Decoder notes the brushwork shows on enlargement. What the number measures is the scale of the program's output, not photographic deception.
No undo is the interesting constraint
Graczyk's version is the stricter test. The model could only place strokes by outputting each one as its own line of coordinates. Once a stroke was down, it stayed down. The bird at sunrise took exactly 111 strokes, starting with a 90-pixel-wide stroke in #30346a and a wobble of about seven pixels that the model said it added because a human arm does not draw straight either.
Under that rule, the useful behavior is not the first stroke. It is the correction. When the bird's back came out too thin, the model laid a bolder stroke over it and the stray bristle marks ended up reading as ruffled feathers. In the 280-stroke piece, a lighting effect smeared objects on a table and the model repainted them. That loop — look at the canvas, name what went wrong, cover it with a better stroke — is the same loop agent demos are judged on in coding. Painting just makes the mistakes visible to anyone, not only to people who read code.
Whose demo is this
Every painting in this story comes from Anthropic staff, working with a model their employer sells, posted on their own accounts. Eaton also steered the work: he had the model invent artist personas first, after earlier attempts with the sibling model Opus 5.5 were criticized for imitating existing art, and he redirected paintings with comments between passes. None of that makes the results fake. It does make them curated. There is no shared task, no blind comparison against a diffusion model given the same prompt and the same time, and no published cost for a 700,000-stroke picture. A reader cannot tell from these posts whether painting this way is a durable capability: employee-selected demonstrations leave performance across other subjects, and the amount of human intervention involved, unresolved.
The one outside test in the report points in both directions. Zhu ran Fable 5.2 and GPT-6 Astra on physical robot arms painting the same Van Gogh. Astra was faster, and OpenAI's model remains the only one to have stood a domino on end in his testing. Zhu's test shows the planning transfers to a robot arm in his setup. It is not yet a lead: the reported corrections illustrate behavior worth testing under repeatable conditions, and they do not by themselves establish general planning reliability.
What a reader should do with it
Treat this as evidence about where language models are heading, not as a new way to make illustrations today. Nobody should replace an image model with a thousand lines of Python on the strength of an employee's thread. The claim worth keeping is narrower: a model that can plan 111 strokes, notice a thin back on a bird, and fix it without an undo key is exercising the same plan-act-inspect-correct discipline that long agent tasks need. Painting is a clean public test of that discipline because the failures are on the surface.
If Anthropic wants the stronger claim, the next step is easy to name: publish the programs, the prompts and the time and token cost for each piece, and let people outside the company run the same subjects. Until then, the honest summary is that the demos are striking, the method is genuinely different from diffusion, and the scoreboard is still being kept by the home team.
Sources
- [1] The Decoder, "Anthropic's Claude Fable 5.2 doesn't generate images, it paints them," November 15, 2026Read source