Back in April, Anthropic surprised the AI world when it revealed Claude Mythos, a model so capable at cybersecurity tasks that the company decided not to release it publicly.
Instead, Anthropic limited access to a small group of organizations through Project Glasswing, a program created in partnership with the U.S. government and focused on cyberdefense and critical infrastructure protection. At the time, the message was clear: some AI capabilities were simply too powerful to put into everyone’s hands.
Now Anthropic is changing course — sort of. The company has launched Claude Fable 5, a new model built on the same underlying system as Mythos 5. But unlike Mythos, Fable 5 is available to the public. The catch is that Anthropic has wrapped it in a series of safeguards designed to prevent users from accessing some of the model’s most powerful capabilities.
In many ways, Claude Fable 5 may be the most important AI release of the year because it signals a new strategy for deploying frontier models.
Making Claude safer for the general public to use
According to Anthropic, Claude Fable 5 and Claude Mythos 5 are fundamentally the same model underneath. The difference is that Fable includes guardrails that intervene when users enter sensitive territory such as cybersecurity, biology, chemistry and certain forms of model extraction. When that happens, requests are redirected to a less capable model instead.
Anthropic says the system works in at least 95% of sessions without needing to fall back to another model, suggesting most users will never notice the restrictions.
The result is effectively a consumer version and a restricted version of the same frontier AI system. That’s new. Historically, AI companies have released a single model and adjusted safety filters around it. Anthropic is now drawing a more explicit line between what the public gets and what trusted organizations receive.
The takeaway
What’s most interesting isn’t Claude Fable 5, but what comes next. For years, AI companies have debated whether increasingly capable models should be released openly, restricted entirely or carefully controlled. Anthropic may have just introduced a fourth option: release the model publicly, but selectively disable the capabilities considered most dangerous.
If that approach works, we may see more frontier AI systems arrive in two forms — a public version for everyday users and a restricted version reserved for governments, researchers and vetted organizations.
Follow Tom’s Guide on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds. Subscribe to Tom’s Guide on YouTube and follow us on TikTok.
More from Tom’s Guide
