Anthropic is already putting its next model into the hands of early access customers, describing it in leaked documents as its most capable yet and warning of “unprecedented” cyber risks.
The AI research lab appears to be sunsetting the Opus name in favour of Mythos, which will serve as its general-purpose model. It will deliver “meaningful advances in reasoning, coding, and cybersecurity,” according to the leaked document, and could be released in the next few weeks.
The leak was first reported by Fortune, which identified nearly 3,000 assets linked to the Anthropic blog that referenced Mythos and Capybara, the company’s next two model releases. Capybara appears to be a new tier, even larger than Mythos, and may only be available to paying customers.
Currently, Anthropic has three tiers of AI models: Opus, Sonnet, and Haiku. Mythos may replace Opus, while Capybara would become a fourth tier focused on software coding, academic work, and cybersecurity.
“Compared to our previous best model, Claude Opus 4.6, Capybara achieves dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity, among others,” Anthropic said in the leaked post.
Both Anthropic and OpenAI warned that their models released last month — Opus 4.6 and GPT-5.3, respectively — show high capability in cybersecurity-related tasks. In the hands of bad actors, these models could be used for a range of attacks, from large-scale phishing campaigns to targeted code vulnerability exploits.
Anthropic also claimed several Chinese AI developers have attempted distillation campaigns to extract its capabilities.
The next stage of AI capabilities
OpenAI and Anthropic are moving at unprecedented speed to release new models, with both reportedly abandoning key safety and ethics pledges to accelerate development. As the race for users, particularly in the enterprise, intensifies, so too do questions around how much risk businesses, governments, and individuals should accept.
Beyond the risks, Anthropic has also taken on OpenAI’s ability to move markets through announcements. Two recent agent-related launches, one targeting finance and another law firms, triggered declines in a tranche of related stocks as investors reacted to potential disruption.
Trying to maintain safety while competing
Both Anthropic and OpenAI face growing competition from tech giants and AI startups, while positioning themselves as research labs committed to safe development.
This tension was evident in Anthropic’s recent fallout with the Pentagon, which sought unrestricted use of its models. Anthropic pushed for limits on applications such as mass surveillance and autonomous weapons, leading to the collapse of a government contract that OpenAI later secured.
Also read: Anthropic has generally positioned itself as more safety-focused, launching an ethical constitution for Claude and a job disruption tracker to monitor where AI is having the greatest impact on employment.