Anthropic has released Claude Fable 5.2 — a new version of its AI model that doesn’t generate images. It paints them. Line by line. Stroke by stroke.
In code. And it’s already solving black hole physics problems, designing circuit boards, and running its own laptop OS. This isn’t diffusion. It’s digital oil on canvas, simulated down to bristle wobble and paint bleed.
The model outputs Python programs, 1,000–2,000 lines long, that drive a virtual paint engine. Each program places hundreds of thousands of individual brushstrokes, no pre-rendered assets, no traced source, no external image model. Just logic, physics simulation, and iterative self-correction. When asked to paint a bird at sunrise, Fable 5.2 began with #30346a, a 90-pixel-wide stroke, and five deliberately wobbling waypoints, because human arms don’t draw straight lines.
Later, it repainted smeared paint tubes after a lighting pass went wrong. It adjusted feather texture by overpainting thin edges. It made two figures hold brushes in outer hands so their shoulders could touch. All in real time, on a blank canvas, with no undo.
Anthropic engineer Michael Graczyk shared a 21-second video showing exactly how the painting builds from 111 strokes up, each coordinate, color, and width generated live. Developer Jake Eaton, who works on AI safety and agent behavior at Anthropic, demonstrated the model’s ability to invent fictional artist personas with consistent styles, a direct response to earlier criticism that sibling model Opus 5.5 merely imitated existing art.
Fable 5.2 is built for long-horizon technical work. Internal tests show it solved two open problems in black hole physics. It designed three working electronic devices with circuit board manufacturer Diode. And it used over 1,000 sub-agents to program its own laptop operating system in eight days.
That same architecture powers Mythos 5.2, a variant reserved for verified security and biology researchers. In robotics, AI researcher Aris Zhu ran side-by-side physical tests: Fable 5.2 and OpenAI’s GPT-6 Astra both controlled robotic arms to paint Van Gogh’s The Starry Night. On Anthropic’s internal RoboDojo Sim benchmark, Fable 5.2 scored 64.9 percent.
Astra scored 29.8 percent. But real-world limits remain. Astra painted faster. Only Astra managed to stand a domino upright, using gravity as a pivot by slowly loosening its gripper.
Fable 5.2 does not ship as a standalone product. It’s not available via Claude.ai or the Claude desktop app. Anthropic says it’s currently deployed internally and with select research partners. No public API, no pricing, no rollout date.
No mention of consumer access, or even whether it will get one. What’s clear is this: Anthropic isn’t chasing pixel-perfect renders. It’s building agents that reason, plan, simulate, and correct, under constraint. And it’s starting with paint.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.