Anthropic has confirmed the existence of a more advanced, unreleased version of its Mythos model family, underscoring just how fast frontier AI capabilities are outpacing the industry’s ability to safely deploy them.
The company announced Claude Mythos Preview on April 7, 2026, describing it as its most capable model to date. What makes this one different from a typical product launch: Anthropic is not letting anyone outside a controlled research program touch it.
A model that can hack your browser before you finish reading this
The standout capability of Mythos Preview is autonomous cybersecurity, and not the defensive kind. The model can independently identify and exploit high-severity zero-day vulnerabilities across major operating systems and browsers.
Because of that risk profile, Anthropic is not offering general access. Instead, the company launched Project Glasswing, a cybersecurity initiative explicitly scoped for defensive applications. Glasswing comes with commitments of up to $100 million in usage credits.
Mythos 5 and the cautious path to broader access
On June 9, 2026, Anthropic released Claude Mythos 5, an upgraded version with enhanced performance across cybersecurity, biology, and healthcare benchmarks.
Alongside Mythos 5, Anthropic introduced Fable 5, a variant built with additional safeguards designed to reduce misuse risk, available under a more restricted access framework than a typical commercial model.
On July 1, 2026, export controls on Mythos 5 were lifted for certain US organizations, marking the first formal step toward broader deployment of a model class that Anthropic had previously kept entirely behind closed doors.
Anthropic has been explicit that no Mythos-class model will see a general public release until stronger safeguards are in place.