Claude Fable 5.2
Anthropic's Claude Fable 5.2 Paints Images With 700K Python Brushstrokes
Anthropic's new model abandons diffusion for a radically different approach — generating photorealistic images by placing 700,000 individual Python‑driven brushstrokes on a virtual canvas. It scored 64.9% on the RoboDojo Sim benchmark, more than doubling GPT‑6 Astra's 29.8%, and its spatial reasoning extends to physical robot arms.
Brushstrokes, Not Diffusion
Anthropic's Claude Fable 5.2 can create photorealistic paintings — not by refining noise like Midjourney or DALL‑E, but by placing hundreds of thousands of individual brushstrokes on a virtual canvas through Python code. According to The Decoder, the model simulates the physical painting process end to end, including bristle texture, how the paint goes on, and mistakes it then has to paint over.
Each artwork is a Python program of 1,000 to 2,000 lines that drives a simulated paint engine. Claude works without an external image model, without pre‑rendered source images, and without tracing over existing art. The approach resembles agentic image editing — a language model picking tools and evaluating its own results.
- No diffusion Abandons the noise‑refinement approach used by Midjourney, DALL‑E, and Stable Diffusion
- Pure Python Each painting is 1,000–2,000 lines of Python that drives a simulated paint engine with physics
- No tracing Works without external image models or pre‑rendered source images — builds from scratch
700,000 Strokes to Photorealism
Anthropic developer 1 the model's capabilities on X. When asked to create a painting as hard to distinguish from a photograph as possible, Fable 5.2 generated a parking garage scene using roughly 700,000 simulated brushstrokes — a shopping cart sitting under fluorescent lights with reflections on the concrete floor.
Anthropic engineer Michael Graczyk shared a separate experiment: a bird at sunrise painted from exactly 111 coordinate sequences. For the very 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 on purpose — because a human arm doesn't draw a perfectly straight line either.
Between passes, the model looked at the canvas and reacted to its own mistakes. When the bird's back came out too thin, the AI laid a bolder stroke over it so the stray bristle marks along the edges now looked like ruffled feathers. In a second piece — two people painting at dusk, 280 strokes — a late lighting effect accidentally smeared objects on a table, and Claude repainted them after the fact.
No Undo Button
Unlike diffusion models that can regenerate iterations, Fable 5.2 commits to every stroke. There is no undo. The model outputs each brushstroke manually as its own coordinate line, with no way to erase strokes once placed. This sets it apart from AI‑generated videos where models like Flux 3 generate image sequences with direct temporal consistency, according to The Decoder.
Earlier attempts with the sibling model Opus 5.5 drew criticism for merely imitating existing art styles. Eaton responded by having Fable 5.2 develop its own artist personas — with fictional backstories and consistent styles — that guide each painting's brushwork through the Python code.
RoboDojo Sim: 65% vs 30%
On Anthropic's internal RoboDojo Sim benchmark, Fable 5.2 scored 64.9% compared with 29.8% for OpenAI's GPT‑6 Astra — a gap of more than 2‑to‑1. The benchmark tests a model's ability to plan step‑by‑step and self‑correct under physical constraints.
AI researcher Aris Zhu tested both models with a physical robot arm painting Van Gogh's The Starry Night The real‑world test exposed limits: while Fable 5.2's spatial reasoning was superior in simulation, Astra painted faster on the physical arm. And so far, only OpenAI's model has managed to stand a domino on end — it did so by slowly loosening its gripper to use gravity as a pivot point.
- RoboDojo Sim Fable 5.2: 64.9% | GPT‑6 Astra: 29.8% — more than double the score
- Physical robot arm Fable 5.2 wins on spatial reasoning, Astra wins on speed and fine motor control
- Domino test Only GPT‑6 Astra has managed to stand a domino on end using a robot gripper
From Canvas to Robotics
Fable 5.2's ability to plan step by step and self‑correct under physical constraints isn't limited to virtual canvases. According to The Decoder, the same spatial reasoning that lets the model paint also transfers to physical robot arms and long‑horizon technical tasks.
In internal tests, Fable 5.2 solved two open problems in black hole physics, designed three working electronic devices with circuit board manufacturer Diode, and used more than 1,000 sub‑agents to program its own laptop operating system in eight days. Anthropic positions the model as built for "extremely demanding long‑horizon technical tasks," The Decoder reports — alongside Mythos 5.2, a version designed for verified security and biology researchers.
What This Means for AI Development
Fable 5.2 represents a shift in how language models interact with the physical world. By generating Python code that simulates real physics — bristles, wet paint, canvas texture — rather than hallucinating pixels, it opens a new path for models that reason spatially instead of statistically.
For builders, the implication is clear: the line between code generation and physical‑world interaction is blurring. A model that can paint 700,000 brushstrokes with physics‑constrained Python can also design circuit boards, control robot arms, and coordinate sub‑agents. The approach is not limited to art — it's a general‑purpose spatial reasoning engine that happens to output paintings as one of its demonstrations.
Sources
- 1.The Decoder(the-decoder.com)
Related News
Nov 16, 2026
Anthropic IPO Targets $1.4 Trillion Valuation, Six Times Oversubscribed Ahead of Nasdaq Debut
Anthropic's IPO is already six times oversubscribed just three days into its roadshow, targeting a $1.4 trillion valuation ahead of its Thursday Nasdaq debut. The listing lands in the middle of an Nvidia leadership crisis and a broader AI stock rout that has erased $4.2 trillion from picks-and-shovels AI firms since June.
Nov 15, 2026
Anthropic Says Your Claude Prompts Are Outdated — Here's What to Change
Anthropic published new prompting guidance for Claude Opus 5.5 telling developers to drop 'think hard' and 'think step by step' from their system prompts. The model's always-on adaptive reasoning makes these legacy instructions redundant — and in some cases actively harmful to speed and cost.
Nov 15, 2026
OpenAI Reorganizes Robotics Team; Robotics Work Continues
OpenAI says it reorganized its robotics team to work more closely with research and product, not disbanded it. The company says its robotics work continues, correcting the shutdown framing in the original report.