Anthropic CEO Dario Amodei has spent much of 2026 cautioning about the existential risks of superhuman AI systems even as his company rolls out increasingly powerful models at a steady clip.

On June 9, 2026, Anthropic launched Claude Fable 5, the first publicly available model in its Mythos class. It arrives with enhanced safeguards designed to blunt the sharper edges of its predecessor, the Claude Mythos Preview, which debuted on April 7-8 and remains locked behind restricted access. The company decided the original Mythos Preview was simply too risky for widespread deployment.

The model too dangerous to ship

Anthropic built the Mythos Preview primarily as a frontier model for cybersecurity. Amodei highlighted its ability to autonomously discover software vulnerabilities, including zero-day exploits. He acknowledged the potential for “destabilizing societal effects” if such a system were widely accessible.

The Mythos Preview was restricted to selected security partners only. No API for developers. No chatbot for consumers. Just a controlled rollout to vetted organizations.

Fable 5 and the art of controlled escalation

Fable 5 represents Anthropic’s attempt to thread the needle. It’s a Mythos-class model, meaning it shares architectural DNA with the restricted Preview, but ships with additional safeguards designed to prevent the most concerning use cases while preserving the performance gains that make the Mythos generation worth using.

The philosopher-CEO problem

In January 2026, Amodei published an essay titled “The Adolescence of Technology,” addressing the risks posed by superhuman AI systems. Anthropic was founded in 2021 by former OpenAI executives who left specifically because they believed AI development needed a stronger emphasis on alignment with human values.

Public sentiment and expert opinion remain split. Some view Anthropic’s approach as the most realistic path forward: develop the technology, understand its risks firsthand, and build safety measures from the inside. Others see it as a convenient justification for racing to build increasingly dangerous systems while wrapping the effort in cautionary language.

What this means for investors

A model capable of autonomously discovering zero-day exploits has obvious commercial applications for defense contractors, enterprise security firms, and government agencies. The restricted partner program for the Mythos Preview suggests Anthropic is already monetizing this capability through high-value, low-volume deals.

The tiered release strategy — restricted access for the most powerful model, public access for the safer variant — could become an industry template. If regulators eventually mandate similar approaches, companies already operating this way will have a structural advantage.