Anthropic has been quietly testing what it calls the most powerful AI model it has ever built—and the only reason the world knows about it is because the company accidentally left the draft announcement in a publicly accessible data cache. The model is called Claude Mythos. It introduces an entirely new tier above the existing Claude lineup, and according to Anthropic’s own draft, it is so capable in cybersecurity that releasing it broadly right now could do more harm than good.When Fortune reached out, Anthropic confirmed the model’s existence. “We’re developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity,” a spokesperson said. “Given the strength of its capabilities, we’re being deliberate about how we release it.” The details first emerged after Fortune reviewed draft blog posts left in a publicly accessible data cache before Anthropic locked it down.
Mythos introduces a tier above Opus that Anthropic has never had before
Mythos isn’t just a more capable version of an existing Claude model. According to the draft reviewed by Fortune, Anthropic is introducing it as the first entry in an entirely new tier it’s calling “Capybara”—sitting above Opus, which was previously the company’s most powerful line. This makes it the first time Anthropic has added a new tier to its model hierarchy, which until now has been structured around Haiku, Sonnet, and Opus.The name, per the leaked post, is meant to evoke “the deep connective tissue that links together knowledge and ideas.” On benchmarks, Mythos significantly outperforms Claude Opus 4.6 across software coding, academic reasoning, and cybersecurity tasks. Anthropic confirmed as much to Fortune, describing it as “a step change and the most capable we’ve built to date” and saying it reflects “meaningful advances in reasoning, coding, and cybersecurity.“
Anthropic says the model is so good at hacking that releasing it would be dangerous
Here’s the catch. The draft reviewed by Fortune states that Mythos is “currently far ahead of any other AI model in cyber capabilities”—capable of discovering and exploiting vulnerabilities in ways that significantly outpace what human security teams can handle. That’s not a third-party assessment. That’s Anthropic’s own evaluation of a model it built.The concern is real enough that Anthropic has structured the entire early rollout around it. Rather than a standard launch, the company is giving select cybersecurity organisations first access—the logic being that defenders need a head start to harden their systems before the model becomes widely available. It’s a similar position to the one OpenAI took earlier this year when it flagged GPT-5.3-Codex as its first model classified as “high capability” for cybersecurity tasks under its Preparedness Framework.
The model is also too costly to run—so a broad release was never close anyway
The restricted rollout isn’t just about safety. The leaked draft also flags that Mythos is a large, compute-intensive model—expensive for Anthropic to serve and, by extension, expensive for customers to use. The company says it’s working to make it more efficient before any general release.So even if the cybersecurity concerns didn’t exist, a wide launch may not have been on the cards regardless. According to the draft reviewed by Fortune—which Anthropic has not formally confirmed in its entirety—the plan is to slowly expand access through the Claude API over the coming weeks, with cybersecurity use cases reportedly getting priority in the early access programme.
