Anthropic just gave the public its first taste of Mythos-class AI, and it arrives with a critical asterisk: the advanced cybersecurity functions that make its unrestricted siblings so powerful have been intentionally limited or removed entirely.
The full Mythos models can autonomously discover and exploit zero-day vulnerabilities — they can find security holes that nobody knows about and break through them without human guidance. Handing that capability to everyone with an API key would be, to put it mildly, inadvisable.
Two models, two very different leashes
Claude Mythos Preview dropped back in April 2026, accessible only to a small group of vetted organizations specifically because of its autonomous zero-day discovery capabilities. That model was essentially a cybersecurity weapon that could probe major operating systems and browsers for exploitable weaknesses.
Now there are two new entrants. Claude Fable 5 is the public-facing product, priced at $10 per million input tokens and $50 per million output tokens. It outperforms Claude Opus on benchmarks, making it the strongest model Anthropic offers to general users. But when queries venture into risky cybersecurity territory, Fable 5 redirects them to older, less capable models like Opus.
