San Francisco, April 25, 2026 — Anthropic has introduced Claude Mythos Preview, its most capable frontier AI model to date. This general-purpose large language model represents a significant advancement in AI capabilities, particularly in agentic coding, complex reasoning, and cybersecurity applications.

 

What is Claude Mythos Preview?

Claude Mythos Preview (internally codenamed “Capybara”) is the latest addition to Anthropic’s Claude family of AI models. It builds on previous releases like Claude Opus 4.6 and demonstrates a notable leap in performance across multiple domains, including software engineering, mathematics, long-context reasoning, and practical research assistance.

Key strengths highlighted by Anthropic include:

  • Advanced agentic capabilities — the ability to autonomously plan, execute multi-step tasks, debug code, and iterate on solutions with minimal human guidance.
  • Strong performance in cybersecurity tasks, where it can identify vulnerabilities in software systems and, in controlled tests, develop proof-of-concept exploits.
  • Improved results on benchmarks such as SWE-bench (coding), USAMO (mathematics), and various internal cybersecurity evaluations.

The model excels at understanding both text and image inputs, making it versatile for conversation, analysis, coding, and creative workflows.

Project Glasswing: Defensive Cybersecurity Initiative

Rather than releasing Mythos Preview for general public use, Anthropic has launched Project Glasswing, a targeted program to harness the model’s capabilities for strengthening digital security.

Under this initiative:

  • The model is being made available to a limited group of trusted partners, including major technology companies such as Amazon, Apple, Microsoft, Google, and cybersecurity firms like CrowdStrike.
  • Anthropic is committing up to $100 million in usage credits and $4 million in direct donations to open-source security projects.
  • The primary goal is to help partners scan critical software infrastructure, identify previously unknown (zero-day) vulnerabilities, and patch them before malicious actors can exploit them.

Anthropic has emphasized that Mythos Preview’s cybersecurity skills represent a “step change,” enabling it to autonomously discover high-severity issues in major operating systems and web browsers during authorized testing.

Responsible Development and Safety Focus

Anthropic’s decision to limit availability reflects its commitment to responsible AI deployment. The company published a detailed System Card outlining the model’s capabilities, safety evaluations, and alignment assessments.

According to Anthropic:

  • Mythos Preview does not yet cross the threshold for fully automated AI research and development.
  • It shows strong alignment properties while acknowledging the dual-use nature of its advanced coding abilities.
  • Future iterations aim to broaden safe access once robust safeguards are in place.

This approach aligns with Anthropic’s long-standing philosophy of prioritizing safety and societal benefit in AI development.

Broader Implications for the AI Industry

Claude Mythos Preview highlights the evolving role of frontier AI models in both advancing technology and addressing its challenges. By channeling powerful capabilities toward defensive cybersecurity, Anthropic is contributing to proactive protection of digital infrastructure in an era of increasing cyber threats.

The model also underscores the rapid progress in AI agentic systems — tools that can independently handle complex, real-world tasks. As the industry moves forward, such focused, responsible deployments may serve as a model for balancing innovation with risk management.

Anthropic continues to iterate on its Claude lineup, with publicly available models like Claude Opus 4.7 offering enhanced performance for general users while more specialized systems like Mythos Preview support critical infrastructure protection.

This article is based on official announcements from Anthropic, including Project Glasswing, the Claude Mythos Preview System Card, and reports from news agencies as of April 25, 2026.