Anthropic just shipped Claude Fable 5. They are calling it a Mythos-class model made safe for general use, and the benchmarks back the claim. Software engineering, knowledge work, vision, scientific reasoning — the longer and more complex the task, the wider the gap between Fable 5 and everything they have previously released.
The more notable part is how they handled safety. Fable 5 is capable enough in areas like cybersecurity, biology, and chemistry that Anthropic decided not to ship it at full power for general use. When your query touches those areas, the model routes it to the older Opus 4.8 instead and tells you it is doing so. That level of transparency about a deliberate capability tradeoff is unusual for an AI lab. They also acknowledge outright that Fable 5 will refuse harmless requests on purpose — less than five percent of sessions in testing, they say — because the company decided that was a reasonable cost. That sounds small until it is your session and you are doing legitimate security research or biotech work.
Then there is Mythos 5, the same underlying model with some safeguards lifted, reserved for a vetted group of cyberdefenders and infrastructure operators through something called Project Glasswing, with US government involvement. Everyone else gets the guardrailed build. Anthropic is running a two-tier system in public: one version for general users, a more capable version for people the government has already cleared.
This tiered access model is going to become standard across the industry. The capability gap between what a general user can access and what a vetted enterprise or government operator can access is already real, and it is only going to widen.