Anthropic leak reveals details of new Claude model.

A CMS error has led to the exposure of approximately 3,000 internal Anthropic documents. The most significant revelation involves a new model currently in development, referred to in drafts as "Mythos" and "Capybara." Company representatives have confirmed that the product is undergoing testing.

Mythos (or Capybara) represents a new class of models, positioned a tier above the current flagship, Opus. Developers are reporting a "qualitative leap" in reasoning, coding, and cybersecurity, with test results significantly outperforming Opus 4.6 benchmarks.

The leaked documents suggest that the model's vulnerability-hunting capabilities could pose a potential security risk. Consequently, the rollout will be extremely cautious, with API access initially limited to a small early-access group. Another hurdle for the release is the high inference cost; Anthropic is reportedly working to optimize the architecture to ensure the model's commercial viability.