Claude Mythos: the “too dangerous” AI model that Anthropic doesn’t want to launch

Anthropic has a new product with a big catch: it’s too powerful to launch.

For a company valued at around US$380 billion and which, according to the market, is preparing to go public later this year, it is an unusual stance — but one that could pay off in the long term.

The new AI model is called Claude Mythos and is the first that Anthropic has publicly acknowledged is too risky to release to the general public. (If the name sounds familiar, it’s probably because you’ve already read about it here, when Fortune revealed blog posts mentioning the model in an open data package on the internet.) Rival OpenAI made a similar decision in 2019, when it held off on releasing GPT-2 out of fear that it would be used to produce very convincing fake texts — at the time, Anthropic’s current CEO, Dario Amodei, was still working with Sam Altman.

This time, Amodei chose another path. The company announced this Tuesday that it will release Mythos by invitation only, within an initiative called Project Glasswing, aimed exclusively at defensive cybersecurity work and restricted to around 40 organizations. The idea is to give “good teams” an advantage in protecting some of the world’s most critical software systems against security risks brought by advanced AI models. Partners include Amazon, Apple, Microsoft and Cisco.

So what does this all mean for Anthropic’s position in the AI ​​race — and the IPO everyone’s hoping for? A few things.

As my colleague Jeremy Kahn recalls, Anthropic is on a strong run. The company achieved annualized revenue of around US$30 billion — a number that implies a 58% jump in revenue in March alone and already exceeds the annualized US$25 billion that OpenAI reported in February. (The comparison is not perfect because each one calculates the “run rate” in a different way, but you can clearly see where the two curves are pointing.)

Now, Anthropic has a model that, according to the company’s own benchmarks, leaves competitors far behind. And he even found a way to get involved with some of the biggest names in corporate technology. All this despite the public fight with the Trump administration and two accidental leaks that gained prominence.

In addition to being a real security initiative, Project Glasswing also works as a great reputation-building movement, in the opinion of Paulo Shakarian, professor of artificial intelligence at Syracuse University.

By creating a tight-knit consortium and working directly with large companies, Anthropic is “taking the lead in mitigating these new risks,” he told Fortune. It is a strategy that, according to Shakarian, “goes very well with chief security officers (CSOs) around the world”. In an industry that relies on the constant exchange of threat information, this type of collaboration tends to score big with corporate clients.

But the Mythos’ new firepower also comes at a price. For Richard Whaling, chief researcher at cybersecurity startup Charlemagne Labs, Anthropic may be concerned not only about misuse, but also about the costs of the model.

“I agree with Anthropic’s concerns about the misuse of Mythos, but I think there is also a resource constraint at play,” he said. “Anthropic hasn’t revealed the size of the Mythos, but has hinted that it is many times larger — and much more expensive — than the Claude Opus. I think it’s likely that the company simply doesn’t have enough GPUs and other computing resources to offer the model on a large scale.”

Translation: Anthropic may have built something that is, at the same time, too dangerous and too expensive to be put on the market for everyone, the way it is today.

No one knows how long Mythos will remain out of reach for ordinary users and most companies. Anthropic says it is already working on “security measures” for the model. In general, AI models become cheaper and easier to operate over time. And there are always customers willing to pay more for extra capacity. Anthropic itself has already announced that it will cover the first US$100 million cost for Glasswing participants, and initial estimates indicate that it could charge these partners around five times more for Mythos than it currently charges for Opus.

OpenAI, of course, is not out of the game. There are reports that it is also about to launch a new model and is preparing a similar rollout for another product with a strong focus on cybersecurity. But for now, Anthropic is in an enviable position in this crazy AI race: with a capability advantage and increasingly locked in to the type of enterprise and security customers it wants to win.