Anthropic leaks internal documents: Claude Mythos is more powerful than Opus, but its release has been delayed due to security risks.

A draft of a suspected internal blog post from Anthropic has been leaked and is quickly spreading within the community. The document reveals that Anthropic is developing a new model tier codenamed “Mythos”—larger and smarter than the existing strongest Opus series—along with a second version called “Capybara.” However, due to the high cybersecurity risks posed by the model’s capabilities, the release plan has been postponed.

Mythos: A New Tier Above Opus

According to the leaked document, Anthropic positions Mythos as “a new tier that is larger and smarter than the Opus models,” explaining that the naming inspiration comes from “the deep structural organization that connects knowledge and ideas.” The document notes that Mythos has achieved “significantly higher” scores in areas such as software development, academic reasoning, and cybersecurity testing compared to the current strongest model, Claude Opus 4.6.

Discussions indicate that Mythos is v1 and Capybara is v2, both versions belonging to the same tier, above Opus.

Release Delayed Due to Cybersecurity Risks

The most striking part of the document is that Anthropic has unusually acknowledged the “significant cybersecurity risks” posed by the model prior to its release. This has been seen by multiple observers as a new signal in the AI industry—marking the first time a leading lab has publicly addressed the risks associated with capability enhancements before a model’s launch, instead of waiting until afterward.

Analysts suggest that Mythos may be the version that Anthropic truly reserves internally, while Capybara is the version released to the public after security processing.

Community Reaction: Mixed Expectation and Concern

After the news broke, the community’s response has been polarized. On one hand, some are excited about the new capabilities; on the other hand, some users worry that a more powerful model means a quicker depletion of token limits and higher subscription fees—some jokingly suggested, “Charge $100 per prompt.”

There are also voices reminding that the document itself leaked through an “unencrypted public database”—which is an ironic cybersecurity twist: a model delayed for being too strong in cybersecurity has its related documents leaked in the least secure manner.

Currently, Anthropic has not issued an official response regarding this matter.

This article “Anthropic Leaks Internal Document: Claude Mythos Stronger than Opus, but Delayed Due to Cybersecurity Risks” first appeared in Chain News ABMedia.