Anthropic has released Claude Fable 5, the public version of a model it once said was too dangerous to ship.

Ethan Mollick spent years describing AI use as casting spells, and now he’s not sure he’s the wizard anymore.

The model ran for nine and a half hours unsupervised and delivered working software he could barely fault.

But the same system hands its most dangerous capabilities to the NSA while the public gets the sanded-down version.

This video asks what it means when the most useful AI is also the least legible.

Sources to Google.