Outlook Business Desk
Anthropic has said its upcoming AI model, ‘Claude Mythos’, outperforms all its earlier large language models, marking a clear leap in capability. According to a Fortune report, the company described it as its most powerful system so far.
The unreleased model came into focus after cybersecurity researchers Roy Paz and Alexandre Pauwels found a draft announcement in an unsecured and publicly searchable cache, which revealed important details about the AI system before its official release.
Anthropic confirmed the incident and said a setup error in its content management system (CMS) allowed the draft blog to become publicly accessible, exposing key details online before any official announcement was made.
The leaked document identifies the system. Anthropic has completed its training and has started a limited rollout by providing early access to a small group of selected users for initial testing.
Mythos is part of a new category of models named Capybara. Earlier, Anthropic structured its offerings into Opus, Sonnet and Haiku tiers, with Opus representing the highest level of capability.
The Capybara tier sits above Opus models in both scale and intelligence. These systems are expected to offer stronger performance but will likely demand far greater computing power and higher costs to operate.
Internal comparisons suggest the new model performs far better than Claude Opus 4.6 in areas such as software development, academic reasoning and cybersecurity, reflecting a notable upgrade in its overall capabilities.
However, Anthropic has flagged concerns that such advanced capabilities could be misused, as the model may detect and exploit system weaknesses more efficiently, potentially enabling large-scale cyber attacks by malicious actors.