Anthropic Tests Claude Mythos: Leak Reveals Model with 10 Trillion Parameters

Anthropic may have accidentally revealed to the market the next level of its models sooner than it was ready. From an open cache of the company, drafts of posts, internal documents, and thousands of unpublished materials were leaked to the network, from which it follows: the new Claude Mythos model is noticeably stronger than the current flagships and may prove too risky for a regular public launch. Based on these data, this is not a cosmetic update to the Claude lineup, but a system that the company internally considers the most powerful AI in its history and therefore launches with extremely limited early access.

Researchers in cybersecurity discovered the leak, finding almost 3,000 files related to the CMS and Anthropic's internal drafts. Among them were drafts of announcements, images, documents, and materials that the company has not yet published. Anthropic confirmed the fact of the leak and acknowledged that customers with early access are already testing Mythos.

According to a company representative, the model provides a qualitative leap in AI capabilities. This is an important detail: this is not just rumors from social networks, but a set of materials whose existence the company did not actually deny, although it did not reveal all characteristics officially. The scale of the discovery makes the story unusual: instead of one randomly surfaced page, the market saw a large body of internal preparations for launch, and before any full public announcement.

If the leaked drafts are to be believed, Claude Mythos should stand above the current hierarchy of Haiku, Sonnet, and Opus. In the documents, a new Capybara tier appears—likely a designation for the most expensive and performant configuration based on Mythos. This model, it is claimed, surpasses Claude Opus 4.

6 in three directions at once: programming, academic reasoning, and cybersecurity. The network also circulated an estimate of 10 trillion parameters, but it is important to understand that this is not yet a confirmed figure from an official release. Even if the parameter count turns out to be less, the overall meaning of the leak remains the same: Anthropic is preparing a model that it considers substantially more powerful and noticeably more expensive to operate.

Hence the caution with the release: such systems require more computation, cost more in inference, and are more difficult to adapt to a regular mass-access model. The main difference of Mythos from familiar launches is the emphasis on risks. In the leaked materials, the model is described as a system that significantly outpaces other AIs in cyber capabilities.

For the market, this sounds like a warning: the next leap in generative models may be associated not only with better coding, but also with more effective vulnerability discovery, exploit building, and attack automation. Earlier, Anthropic had already spoken about cases where its tools were attempted to be used in malicious campaigns, and Opus-level models have learned to find previously unknown vulnerabilities in real code. Therefore, Anthropic's strategy looks unusual: first give access to organizations associated with defense, and only then decide on a wider release.

This approach fits into the overall logic of recent months, when frontier model developers increasingly discuss not only benchmarks, but also the danger threshold, after which a model's capabilities become a problem in themselves. For developers and companies, two conclusions are important here. The first: the race for models is shifting from competition in "who writes text better" to competition in "who works better with code, infrastructure, and security."

The second: access to the most powerful systems will increasingly be limited not by marketing, but by real considerations of cost, control, and prevention of abuse. Claude Mythos may turn out to be another exaggerated leak, and some of the figures and formulations will certainly be revised. But the trend itself is already obvious: Anthropic is testing a new level of AI and doing so with maximum caution.

If the company truly sees in the model a tool capable of simultaneously helping defenders and empowering attackers, the mass release of such systems ceases to be a technical question and becomes a question of access policy.