Claude Mythos Preview — Anthropic's new frontier, still closed. Launching as part of Project Glasswing: an initiative with AWS, Google, Microsoft, Apple, Cisco, CrowdStrike, NVIDIA, and several other companies. The goal is cybersecurity.
Over a few weeks, the model autonomously found thousands of zero-day vulnerabilities — in every major OS and every major browser. Three examples: a 27-year-old vulnerability in OpenBSD, a bug in FFmpeg that automated tests ran over five million times without catching, and a chain of holes in the Linux kernel giving full control over a machine.
A month ago Anthropic published a case study with Firefox — back then it was still Opus 4.6, the previous model. It found 22 vulnerabilities in two weeks, 14 of which Mozilla classified as high severity — nearly a fifth of all serious Firefox vulnerabilities patched throughout all of 2025. That was a warm-up. Mythos is a different level.
On benchmarks: SWE-bench Verified — 93.9% (Opus 4.6: 80.8%), GPQA Diamond — 94.6%, Humanity's Last Exam without tools — 56.8%.
Right now only 12 major partners have access to Mythos (partially closed list). Maintainers of significant open source projects can apply separately. Everyone else waits for the next Claude Opus, which will be released publicly — already with built-in safety mechanisms.
After exhausting the allocated $100M in credits, participants pay $25/$125 per million tokens. Sounds expensive. But according to the AI model pricing research (96 models, 6 benchmarks) — this is how every top release launches, and a year later it all costs several times less.