---
title: "Claude Fable 5.2 makes images by coding simulated brushstrokes"
url: https://www.parallelquant.com/posts/claude-fable-5-2-makes-images-by-coding-simulated-brushstrokes-123484
source_name: "The Decoder"
source_url: https://the-decoder.com/anthropics-claude-fable-5-2-doesnt-generate-images-it-paints-them/
published: 2026-11-15T15:42:03.000Z
topics: ["llms", "products"]
publisher: "Parallel Quant"
---

# Claude Fable 5.2 makes images by coding simulated brushstrokes

*2026-11-15 · Source: [The Decoder](https://the-decoder.com/anthropics-claude-fable-5-2-doesnt-generate-images-it-paints-them/)*

Anthropic's Claude Fable 5.2 creates images without generating pixels directly. Instead, the language model writes Python code that places hundreds of thousands of simulated brushstrokes, with bristle texture, on a virtual canvas. Anthropic employee Jake Eaton said one photorealistic parking-garage scene took about 700,000 strokes.

**Why it matters:** This shows a text-only model getting visual output by writing and running programs, not through a built-in image decoder. The same idea, using code as the model's hands, drives today's agent work. Because the output is a procedure rather than a fixed bitmap, it can be edited, inspected and re-rendered in a way diffusion images cannot. It also says something about the spatial reasoning these models have built up, coming as Fable 5.2 tops math leaderboards.

**Topics:** llms, products

---
Read the original: https://the-decoder.com/anthropics-claude-fable-5-2-doesnt-generate-images-it-paints-them/
Canonical: https://www.parallelquant.com/posts/claude-fable-5-2-makes-images-by-coding-simulated-brushstrokes-123484
Published by Parallel Quant — https://www.parallelquant.com
