Claude Fable 5 generates full games from a single prompt

This week, Anthropic launched their latest flagship model, Claude Fable 5 (the first model in their new Mythos line), and independent assessments are looking good. AI researcher Ethan Mollick of UPenn was able to create functional video games using only one prompt per game.

On his to-do list? Classic Snake, a cave exploration game called Strata, and a poetry game called Duino based on Rilke’s poems. He also had the model generate an interactive time visualization tool in response to one prompt. “Fable outperformed basically every other public model I have used by a considerable margin,” he said. And he found that the model was able to run through complex, multi-page specifications for up to a dozen hours straight.

1Why the long runs matter

At the end of the day, autonomous runtime is getting longer. Last year, if your model could complete a task autonomously within twenty minutes without crashing, you were doing well. Today, leading edge models are holding onto long, complex plans for hours. As a non-coder, this means going from “describe a tool and then baby-sit it” to “describe a tool and then go live your life while it gets done.”

Games are cool, but what this really shows is the ability to convert natural language descriptions into fully functioning software, with automated tests, just like you would ask an AI code generator to create a budgeting app or a reservation system. If a model can handle an entire game, five screens of your personal utility should be a warmup for it.

A word of warning, though. Just because models are running autonomously for longer doesn’t mean they’re running correctly. You still have to check their output and test it yourself before you assume it’s working. That will always be true, no matter how advanced the models get.