Anthropic is testing Claude Mythos: a leak reveals a model with a reported 10 trillion parameters

Anthropic may have accidentally shown the market its next tier of models before it was ready to do so itself. Draft posts, internal documents and thousands of unpublished assets spilled from the company's publicly accessible cache. They indicate that the new Claude Mythos model is markedly stronger than today's flagships and may be too risky for an ordinary public launch. According to this data, it is not a cosmetic update to the Claude lineup but a system the company internally regards as the most powerful AI in its history, which is why it is being rolled out only in extremely limited early access.

How the leak was discovered

Cybersecurity researchers found the leak, turning up nearly 3,000 files tied to Anthropic's CMS and internal staging materials. They included draft announcements, images, documents and other material the company had not yet published. Anthropic confirmed the leak and acknowledged that early-access customers are already testing Mythos.

A company spokesperson said the model delivers a step change in AI capabilities. That detail matters: this is not social media rumor but a set of materials whose existence the company has effectively not denied, even though it has not officially disclosed all of the model's specifications. The scale of the find makes the story unusual: instead of a single page surfacing by accident, the market saw a large body of internal launch preparation, before any full public announcement.

What we know about Mythos

If the leaked drafts are accurate, Claude Mythos is meant to sit above the current Haiku, Sonnet and Opus hierarchy. The documents mention a new tier called Capybara, probably the label for the most expensive and highest-performing configuration built on Mythos. That model is said to outperform Claude Opus 4.

6 in three areas at once: programming, academic reasoning and cybersecurity. An estimate of 10 trillion parameters has also circulated online, but it is important to understand that this figure is not yet confirmed by an official release. Even if the parameter count turns out to be lower, the gist of the leak is the same: Anthropic is preparing a model it considers substantially stronger and noticeably more expensive to run.

Hence the caution over release: such systems need more compute, cost more at inference and fit poorly into the usual mass-access model.

A focus on risks and cyber capabilities

The main difference between Mythos and typical launches is the emphasis on risk. The leaked materials describe the model as clearly ahead of other AI systems in cyber capabilities. For the market, this reads as a warning: the next leap in generative models may bring not only better coding but also more effective vulnerability discovery, exploit development and attack automation.

Anthropic has previously described cases in which its tools were used in malicious campaigns, and Opus-class models have learned to find previously unknown vulnerabilities in real code. That makes Anthropic's strategy look unusual: first give access to organizations involved in defense, and only then decide on a wider release. The approach fits the broader logic of recent months, as frontier model developers increasingly discuss not just benchmarks but the danger threshold beyond which a model's capabilities become a problem in themselves.

Takeaways for the market

Two conclusions matter for developers and companies. First, the model race is shifting from a contest over who writes text best to a contest over who works best with code, infrastructure and security. Second, access to the strongest systems will increasingly be limited not by marketing but by real concerns about cost, control and abuse prevention.

Claude Mythos could turn out to be another overblown leak, and some of the figures and wording will almost certainly be revised. But the trend is already clear: Anthropic is testing a new tier of AI and doing so with great caution. If the company really sees the model as a tool that can both help defenders and strengthen attackers, a mass release of such systems stops being a technical question and becomes a question of access policy.

In brief:

• Nearly 3,000 files, including drafts about Claude Mythos, leaked from Anthropic's publicly accessible cache.
• Mythos is described as stronger than Opus 4.6 in programming, reasoning and cybersecurity; the 10 trillion parameters are not officially confirmed.
• Because of risk and cost, the release starts with limited access, primarily for defense organizations.