Claude Fable 5 Built a 3D Browser Game From a Single Prompt

Developer Simon Willison fed Anthropic's Claude Fable 5 one prompt, a handful of screenshots, and some old concept art. What came back was a playable 3D browser game called Raccoon Heist, complete with mobile-friendly touch controls.

Willison ran the model inside Claude Code for web, the coding-focused interface Anthropic has built around its Claude family. The instruction was straightforward: build a 3D game for the browser that works on touchscreens. Claude Fable 5 handled the rest, translating static images and a short brief into working game code.

The project has an unusual backstory. Raccoon Heist's premise dates back to August 2022, when Willison originally generated the game concept using GPT-3 and produced accompanying artwork with DALL-E. That earlier experiment never became a real game. Four years later, a newer model finally turned the idea into something people can actually play.

From concept art to playable code

Beyond writing the game's logic, Claude Fable 5 also generated in-game textures by calling OpenAI's image generation API, mixing Anthropic's coding model with OpenAI's image tools in a single pipeline. That detail underscores a trend already visible across the AI coding space: developers increasingly treat different vendors' models as interchangeable components rather than locking into one ecosystem.

The finished game is live and playable directly in a browser, deployed through GitHub Pages rather than a custom backend or app store submission. That deployment choice keeps the project lightweight and easy for others to inspect, fork, or extend.

Willison's writeup frames Raccoon Heist less as a polished product and more as a demonstration of how far prompt-to-game generation has come since 2022.

What makes the case notable isn't the game itself, a browser-based heist title starring a raccoon, but the compression of effort involved. A task that once required manual coding, asset production, and integration work was largely automated from a single natural-language instruction plus reference imagery.

For developers watching the trajectory of coding-focused AI models, the experiment offers a concrete data point rather than a marketing claim. It shows a model handling 3D rendering, touch input logic, and cross-vendor API calls in one pass, based on unstructured visual references rather than a detailed technical specification.

Anthropic has not issued a formal announcement tied to this specific demo. The details come from Willison's own published account of building and deploying the game, making it a hands-on field report rather than a vendor-driven launch story.