Claude Mythos: The AI Model Anthropic Won't Release — And Why That's the Right Decision

April 13, 2026 – 9 min read

There are moments when you realize we're at a genuine turning point. The story of Claude Mythos is one of those moments.

Anthropic has developed a model — and then decided not to release it. Not because it doesn't work. Because it works too well.

That's new. And it's worth looking more closely.

How Mythos Came to Light

It didn't begin with an official announcement, but with a mistake. In early 2026, a security oversight at Anthropic allowed internal information about a new model to become public — a model with capabilities that no one had on their radar at the time.

Anthropic responded quickly and confirmed the existence of Claude Mythos Preview. The official communication followed in April 2026 — not in the form of a classic product launch, but as an explanation: why this model is not being made publicly available, and what is planned for it instead.

This transparency is remarkable. Other companies would have buried the whole thing. Anthropic made it a matter of principle.

What Mythos Can Do — And What That Means

The capabilities of Claude Mythos Preview are particularly extraordinary in one area: cybersecurity.

That might sound technical and abstract. It's not. Here are concrete facts:

Mythos identified thousands of critical security vulnerabilities — including zero-day flaws in all major operating systems and web browsers

It discovered a 27-year-old bug in OpenBSD — a vulnerability that had gone undetected for nearly three decades

On expert-level CTF (Capture the Flag) tasks — security challenges that no previous AI model had been able to solve — Mythos achieves a 73% success rate

It was the first model to fully solve "The Last Ones" — a 32-step simulation of a corporate network attack

And: it found real, previously unknown zero-day vulnerabilities in Firefox — not simulated, but in production software

On SWE-bench verified, Mythos reaches 93.9% — for comparison: Opus 4.6, the best publicly available model, achieves 80.8%.

This isn't incremental progress. It's a quality leap.

Why It's Not Being Released Publicly

The decision not to release Mythos is not a market strategy. It's a safety decision.

A model that finds zero-days in production software, that can simulate attacking corporate networks, and that can autonomously develop exploits for known CVEs — giving this to the general public would be irresponsible.

That's Anthropic's conclusion. And I think they're right.

Frontier AI development always faces the dilemma between openness and responsibility. Mythos is a case where the balance clearly lies on the side of responsibility. The ability to find real vulnerabilities in widely-used software has too much potential for harm if made available without controls.

Project Glasswing: Responsible Deployment

Instead of a public release, Anthropic has launched Project Glasswing: a consortium of leading technology companies that work together with Mythos Preview — with the explicit goal of identifying and closing security vulnerabilities in critical software before they can be exploited.

The consortium includes:

AWS, Apple, Cisco, CrowdStrike

Google, JPMorgan Chase, Linux Foundation

Microsoft, NVIDIA, Palo Alto Networks

Anthropic is committing $100 million in usage credits to the effort.

This is a genuine paradigm shift. Instead of "release and see what happens," frontier AI is being deployed as infrastructure for proactive security research. The companies most at risk from security vulnerabilities get access to the most powerful tool for finding those vulnerabilities — before attackers do.

What This Means for the Future of AI Development

Mythos is a watershed moment — not just because of its capabilities, but because of the decision that came with it.

It shows that Anthropic is willing to not release a commercially valuable product when the risks are too high. That's a statement about values, not just about product strategy.

For the AI industry as a whole, the question now is: how do other companies handle similar capabilities? Is this the beginning of a new standard — deploying frontier models in controlled environments before making them more broadly available?

I hope so.

Conclusion: Sometimes Restraint Is Strength

Claude Mythos Preview is technically the most impressive AI model I've ever written about. And at the same time, the decision not to release it is equally impressive.

It takes courage to hold back a technology — especially in a market where speed and announcements matter. With Mythos and Project Glasswing, Anthropic has shown that responsible AI development isn't empty words, but lived practice.

That's the state of things in April 2026. And I'm curious what Project Glasswing will discover in the months to come.

Patrik Germann