Anthropic Shipped Two Versions of the Same AI. One for You. One for the Government.

For over two months, Anthropic sat on its most powerful AI model. Not because it wasn't ready. Because it was too ready. Claude Mythos — the model the company calls the strongest cybersecurity AI in the world — was so capable that releasing it publicly would've been handing the internet a master key.

Today they found a workaround. They released Claude Fable 5: Mythos with the dangerous parts lobotomized. Ask it to exploit a cybersecurity vulnerability, develop a bioweapon, or train a competing model — and it quietly hands you off to Claude Opus 4.8, a less capable model that can't complete the task. Three 'classifiers.' Three categories of capability that got quietly rerouted.

The full Mythos 5 goes to a different list. A 'small group of cyberdefenders and infrastructure providers' — curated with U.S. government involvement. Same underlying model. Two tiers of access. A $965 billion company, nine days after confidentially filing for IPO, just told you who gets the real AI and who gets the safe version. Zero independent regulators were involved.