While the world is talking about Opus 4.8, there’s another model in the room that Anthropic is much less willing to discuss: Claude Mythos. It’s Anthropic’s most powerful language model — and it’s not meant for the public. At least not yet.
What is Mythos?
Mythos is part of Project Glasswing, an initiative Anthropic launched in April together with AWS, Apple, Google, Microsoft, NVIDIA, and other major tech companies. The goal: securing critical software infrastructure. Mythos brings advanced cybersecurity capabilities that have raised eyebrows even within Anthropic.
So far, only a handful of organizations have been given access to the Mythos Preview. The reason: the model could be misused by hackers. Anthropic’s engineers are working on specialized guardrails to prevent that.
Why not just release it?
That’s the key question. And Anthropic’s answer is unusually honest: because they’re not sure it’s safe enough. Mythos has cybersecurity capabilities that concern executives and government officials worldwide. Before the model goes wide, protective mechanisms need to be in place.
Anthropic says the guardrails should be ready soon and that Mythos will then be made available to all customers. There’s no specific timeline.
The Glasswing connection
It’s no coincidence that Mythos runs under the Glasswing umbrella. The project brings together the world’s biggest tech companies with a clear goal: using AI models to defend critical infrastructure. On May 22, Anthropic published an initial progress update. The message: things are moving forward, but carefully.
My take
The fact that Anthropic is holding back a model because it might be too capable is remarkable. In an industry where every advantage is typically marketed immediately, this is an unusual move. Whether it’s a PR play or genuine caution — we won’t be able to judge until Mythos actually ships.
What does convince me: Anthropic isn’t just talking about responsible AI, they’re acting on it. Not releasing a model because you haven’t contained the risks yet costs money and market share. The fact that they’re doing it anyway speaks to the credibility of their Responsible Scaling approach.
Sources: