Anthropic Unveils Claude Mythos Preview via Project Glasswing

Anthropic introduced Claude Mythos Preview on April 7, 2026, as part of Project Glasswing, a cybersecurity initiative providing gated access to this frontier AI model. The model excels in cybersecurity, software coding, complex reasoning, and autonomous agent tasks, with a 1M token context window and 128K max output tokens. It outperforms prior models like Claude Opus 4.6 on benchmarks including 93.9% on SWE-bench Verified, 77.8% on SWE-bench Pro, and demonstrates superior vulnerability detection in real-world software such as Firefox, FFmpeg, and major operating systems. Access is restricted to partners like AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, plus 40 organizations maintaining critical infrastructure. Anthropic committed $100 million in usage credits and $4 million in donations to open-source security groups like OpenSSF and Apache Software Foundation. Pricing stands at $25 per million input tokens and $125 per million output tokens via Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft platforms.

Project Glasswing prioritizes defensive use to identify and patch zero-day vulnerabilities, including subtle bugs up to 27 years old, before attackers exploit them. The model autonomously finds and exploits flaws in hardened systems, chaining vulnerabilities in enterprise simulations that take human experts over ten hours. This addresses the dual-use risk of AI capabilities that enhance both defense and offense, with open-source software as a primary focus due to its role in enterprise and critical systems. Availability remains in gated research preview, limited to allow-listed organizations in regions like US East (N. Virginia) on Amazon Bedrock, with no plans for general release to mitigate cybersecurity threats.