A configuration error in Anthropic PBC’s content management system has revealed that it’s testing a new large language model called Claude Mythos.

The company confirmed the project in a Thursday statement to Fortune. According to the company, its engineers have finished training Claude Mythos and are piloting it with early customers. Anthropic added that the LLM is “the most capable we’ve built to date.”

Mythos came to light after the company accidentally left a CMS folder with 3,000 assets publicly accessible. The repository contained a draft version of a launch blog post for Claude Mythos. According to Fortune, the document indicates that the model will be pricier than the company’s existing algorithms.

The blog post also revealed that Anthropic plans to change how it distributes its LLMs. Claude 4.6, the company’s current flagship model, is available in three versions with different feature sets and pricing. Anthropic will add a fourth product tier when it launches Claude Mythos.

The new LLM edition will reportedly be pricier and more capable than Opus, Anthropic’s current top-end product tier. The draft blog post refers to the upcoming offering as Capybara. Anthropic’s internal testing reportedly indicates that the Capybara edition of Mythos is “dramatically” better than Claude 4.6 Opus at programming tasks and other reasoning use cases.

Capybara is particularly adept at finding cybersecurity vulnerabilities. As a result, Anthropic will take precautions to prevent hackers from gaining access to the model.

The model “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders,” reads the draft post. “We’re releasing it in early access to organizations, giving them a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits.”

Shares of CrowdStrike Holdings Inc., Palo Alto Networks Inc. and other major cybersecurity providers dropped more than 5% on the news. Investors might be concerned that Capybara could give Anthropic a competitive edge in the vulnerability detection market. The company entered the segment last month by releasing a tool called Claude Code Security.

The tool is currently powered by Claude Opus 4.6, which has proven highly adept at discovering vulnerabilities. Anthropic engineers have used it to find more than 500 high-severity exploits in open-source projects. On one occasion, the model deduced the presence of a flaw in a PDF tool by analyzing a developer comment in its change log.

The disclosure of Claude Mythos comes a few days after word emerged that OpenAI Group PBC has finished pretraining its newest LLM. Pretraining is the phase of the development workflow in which engineers build a model’s core capabilities. It’s followed by smaller optimizations that focus on improving the LLM’s hardware-efficiency, safety and usability.

OpenAI’s new model is reportedly known as Spud internally. The company is expected to launch it within weeks.

Image: Anthropic