Claude Fable 5 Builds Whole Games From One Prompt

Anthropic released Claude Fable 5 this week — the first publicly available model from the new Mythos family — and the independent results are striking. AI researcher Ethan Mollick of the University of Pennsylvania got working video games out of it, each generated from a single prompt.

His test list reads like a tiny arcade: a retro Snake, a tunnel-exploration game called Strata, and Duino — a poetry game built around Rilke's verses. Beyond games, the model produced an interactive travel-time map from one request. Mollick wrote that Fable "outperformed basically every other public model I have used by a considerable margin", and reported it executing detailed multi-page specifications for up to a dozen hours at a stretch.

1Why the long runs matter

Strip away the benchmarks and one fact remains: the length of an unattended run keeps growing. A year ago, an agent that built something for twenty minutes without stalling was impressive. Now a frontier model holds onto a long, detailed plan for hours. For a non-programmer, that's the difference between 'describe a tool and babysit it' and 'describe a tool, go live your life, and come back to a finished thing'.

The games angle is fun, but it's really a demo of the same skill behind asking an AI coding tool for a budget tracker or a booking page: turning a plain-language description into working software, end to end, including testing its own work along the way. If a model can keep a whole game coherent, your five-screen personal tool is a light jog for it.

One honest caveat: longer autonomous runs don't mean infallible runs. The verdict on whether a thing works is still yours — open what it built and click around before you celebrate. That habit outlives every model upgrade.