Anthropic’s ‘Mythos’ Model Leaked: Claims to Outperform Claude Opus 4.6 Across the Board

A recent configuration error in Anthropic’s public‑facing content‑management system inadvertently exposed about 3,000 internal files, including a draft blog post that announced a new large language model called Claude Mythos (code‑named Capybara).

Mythos is described as a “top‑tier” model that is larger and more intelligent than the current flagship Claude Opus 4.6. Internal tests reportedly show substantial score improvements over Opus 4.6 in software coding, academic reasoning, and cybersecurity benchmarks, with a parameter scale that may reach up to 10 trillion.

The draft also warns that Mythos’s extraordinary capabilities could be weaponized for large‑scale, high‑impact network attacks, making it “far beyond the defensive capacity of current security teams.” Because of this risk, Anthropic has not publicly released the model.

Anthropic’s response has been to shut down the exposed data retrieval channel, acknowledge the human error, and plan to grant early‑access trials to security‑focused institutions so they can prepare for an upcoming wave of AI‑driven vulnerability exploitation. The leak also included a confidential PDF outlining a private CEO summit in the UK, where Dario Amodei will present the model’s undisclosed abilities to invited business leaders.

Beyond the core announcement, many of the leaked files were old marketing assets (images, banners, logos), while a few were clearly private internal documents (e.g., a staff parental‑leave note). The full draft blog post has been archived by community members for reference. (Fortune 2026; Mario Nawfal, X)