Leaked Anthropic materials have offered an unusually detailed look at an unreleased AI model that the company has described as its most capable system so far. The documents refer to the model as Claude Mythos in one version, while another version uses the name Capybara.
The leak matters because the drafts do more than hint at a routine product update. They describe a model positioned above Anthropic’s Opus line, with stronger performance in areas that include software coding, academic reasoning, and cybersecurity.
What the leaked documents say
The exposed materials include two versions of what appears to be the same blog post. One version uses the name "Mythos" and another uses "Capybara." The source says the drafts differ only in the model’s name, which suggests Anthropic may have been weighing different name candidates for the same model.
That impression is reinforced by a mismatch in the Capybara draft. Although the name was changed through the title and body text, the subtitle still says, "We have finished training a new AI model: Claude Mythos." Both versions also explain the name with the same wording, saying it was chosen to evoke "the deep connective tissue that links together knowledge and ideas."
According to the drafts, the model is intended to sit above Anthropic’s existing Opus models. The documents describe it as "larger and more intelligent than our Opus models—which were, until now, our most powerful," and say it achieves "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity, among others" compared to Claude Opus 4.6.
Anthropic’s public response
Anthropic confirmed to Fortune that it is training and testing a new model. The company did not simply dismiss the leaked materials as fabricated. Instead, it characterized the documents as "early drafts of content that were being considered for publication."
A spokesperson also gave Fortune a broader description of the model’s direction: "We're developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity," the spokesperson said. "We consider this model a step change and the most capable we've built to date."
Those comments align with the central claims in the drafts while leaving some important questions unanswered. The final name is still unclear. The timing of any broader release is also uncertain. The documents mention Mythos and Capybara, but the model could still ship under a different name.
Why cybersecurity is central to the story
The most consequential part of the leak may be the way the drafts discuss cybersecurity. According to the documents, Anthropic is planning a deliberately slower release than it has used for previous models. That slower approach is tied to the model’s alleged cyber capabilities.
The drafts say the model is "currently far ahead of any other AI model in cyber capabilities." They also warn that it "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."
That framing explains why Anthropic appears to be taking a cautious rollout path. The source says access will begin with a small group of early-access customers who are tasked with evaluating cybersecurity applications. Access through the Claude API would then expand gradually.
In practical terms, Anthropic is presenting the model as both powerful and sensitive. Stronger cybersecurity performance could help defenders test systems, analyze weaknesses, and evaluate risks. The same capability, however, is why the company’s own draft language points to a need for a slower, security-focused release.
The cost problem
The leaked drafts also describe a major cost issue. The model is said to be "very expensive for us to serve, and will be very expensive for our customers to use." That is a notable constraint for any general purpose AI model, especially one that may eventually be offered through the Claude API.
Anthropic says it is working to make the model "much more efficient before any general release." That statement suggests the company may not see the current version as ready for broad customer access, even if it is already being tested with select users.
Cost and availability are closely linked here. A model can be technically impressive but still difficult to release widely if serving it is expensive. The drafts suggest Anthropic is trying to improve efficiency before making any wider move.
How the leak happened
The documents became public because of a misconfiguration in Anthropic’s content management system. According to the source, a default setting automatically made uploaded files public. As a result, nearly 3,000 internal documents were exposed for anyone to see.
That detail gives the story a second layer. The leak is not only about a frontier AI model. It also shows how internal publication workflows can create risk when default settings expose files that were not meant to be public.
The timing is also notable because OpenAI is reportedly preparing a major release of its own. The source says OpenAI is working on a new model codenamed "Spud," which has already finished pretraining. OpenAI CEO Sam Altman has internally promised a large capability jump, saying the model can "really accelerate the economy."
For now, the clearest facts are narrow but significant: Anthropic is testing a new general purpose model, leaked drafts call it Claude Mythos in one version and Capybara in another, and the company says it represents meaningful advances in reasoning, coding, and cybersecurity. Whether the model reaches customers as Mythos, Capybara, or something else remains unresolved.