How Claude Fable 5.2 paints pictures by writing code for a simulated brush instead of generating pixels

AI & Compute

Anthropic's Fable 5.2 Paints by Writing Code, Not Pixels

Claude Fable 5.2 writes 1,000-2,000 line Python programs that paint stroke by stroke — a 111-stroke bird, a 700,000-stroke shopping cart, and a robotics pitch worth doubting.

By
Nathan Brooks
Filed
Channel
AI & Compute
Read
5 min read

Anthropic released Claude Fable 5.2 on Nov. 9, and its most talked-about trick contains no image model at all: the model writes Python programs of one to two thousand lines that place every stroke through a simulated paint engine. Anthropic researcher Jake Eaton said the model "does not use an image model, nor render or trace in advance." The Decoder counted 111 strokes in a 21-second video of the model painting a bird at sunrise.

That is the hardware of the argument. Where a diffusion model turns noise into pixels over many steps, Fable 5.2 produces a list of brush actions, and every mark is a decision the model can inspect, revise or paint over. The approach trades speed for a record of process. An image model returns a picture in seconds; the photorealistic attempts here run to hundreds of thousands of strokes.

What the model actually does

The model has two jobs, according to Eaton and a Nov. 15 report from The Decoder that collected the staff demos posted on X since launch day. First, it writes or revises the paint engine itself — the code that simulates how bristles deposit paint, how wet paint behaves and how canvas texture shows through. Then it decides each stroke: color, width and the path the brush follows.

Eaton said he steers the process with comments and Claude updates the engines that power the brushes and paint. Most paintings take a few thousand strokes. He began the project after critics said paintings made by the earlier Opus 5.5 merely imitated real art; his response was to ask Fable 5.2 to invent artist personas, each with a back story and a consistent style, and to paint as them.

The most demanding result is instructive. Eaton's most photorealistic attempt, a shopping cart in a parking garage, took about 700,000 simulated strokes, The Decoder reported. The model earns realism one coordinate at a time.

What does the 111-stroke bird demonstrate?

The clearest demo came on launch day from Michael Graczyk, identified by The Decoder as an Anthropic engineer. His constraints were strict. Code could only simulate bristles, wet paint and canvas physics; the model had to output every stroke as its own line of coordinates. "In this setup Claude cannot undo or change anything, it paints like a human," Graczyk wrote.

The chat log, as recounted by The Decoder, shows the model reasoning like a painter rather than a renderer. The first stroke was dark blue (#30346a), 90 pixels wide, with five waypoints and a deliberate wobble of about seven pixels — because, the model explained, a human arm does not draw a perfectly straight line. Between passes it examined the canvas. When the bird's back looked thin, it laid a bolder stroke over it rather than erasing. The finished bird took 111 strokes.

That is the commercial point, thin as it may seem. Each stroke is an explicit choice of position, color and pressure, so the model can review what it has done, plan the next layer and correct mistakes by painting over them. The method is slower and less faithful than diffusion. It is also legible in a way pixels are not.

One caveat belongs here in bold: Anthropic's launch page does not mention painting at all. These are staff experiments, not product features.

Why the robot arm matters

Turning a goal into a series of precise physical actions is the core problem in robotics, and that is where Anthropic is pitching Fable 5.2 in earnest. On RoboDojo Sim, a benchmark in which a model controls a simulated two-arm robot from camera images, Anthropic reports Fable 5.2 scored 64.9% at maximum effort. The company's own figures put Opus 5.5 at 39.4%, Google's Gemini 4 Argon at 34.2% and OpenAI's GPT-6 Astra at 29.8%.

Anthropic ran those tests itself. Its launch video shows robots controlled by Fable 5.2 solving a puzzle cube, unscrewing a bottle cap and drawing a picture of a robot.

An outside test was less one-sided. AI researcher Aris Zhu posted a video of Fable 5.2 and GPT-6 Astra each driving a real robot arm to paint Van Gogh's "The Starry Night." The Decoder, describing the comparison, said Astra painted faster and showed finer motor control, and that only Astra has so far managed to stand a domino on end. One informal test is not a benchmark. All the headline numbers here are staff demos and vendor-run evaluations.

Where this fits in the model's other claims

The paintings join a run of long, multi-step tasks shown with the model. Anthropic said Fable 5.2 designed three working electronic prototypes from schematic to enclosure. It leads the Remote Labor Index of real freelance jobs at 41.3%. Outside the company, logician Jan von Plato used Claude to work through Kurt Gödel's shorthand notebooks.

Pricing is conventional for the frontier tier: $8 per million input tokens and $40 per million output tokens. Anthropic has not said what a single painting costs to produce, and the omission is telling — at a few thousand strokes per picture and 700,000 for the ambitious ones, the token bill is presumably not a selling point.

The open question is whether stroke-by-stroke planning transfers from a simulated canvas to physical manipulation at scale. RoboDojo Sim says yes; a real robot arm painting Van Gogh, slower than a rival, says not yet.

Source: mlq.ai

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

970 articles

Related articles

« Previous article