Anthropic built its most powerful AI model: Mythos; then decides to hold off

Claude Mythos Preview finds thousands of zero-day vulnerabilities, scores 93.9% on SWE-bench Verified, and gets assessed by a clinical psychiatrist. Here's what its 243-page system card reveals.

If you find this kind of deep analysis useful, consider subscribing for more pieces on AI breakthroughs, security, and the frontier of what machines can do.

Today, Anthropic did something unusual for a company in the AI arms race. It announced its most capable model to date, then told the world it wouldn't be releasing it.

Claude Mythos Preview (used in "Project Glasswing"), named after the glasswing butterfly whose transparent wings let it hide in plain sight, represents what Anthropic describes as "a striking leap" over Claude Opus 4.6 across nearly every benchmark worth tracking. On SWE-bench Verified, it scores 93.9%. On USAMO 2026, the proof-based mathematical olympiad, it hits 97.6%. It solves every challenge on the Cybench cybersecurity benchmark with a 100% pass rate.

And you can't use it.

Instead of shipping it to developers and consumers, Anthropic channelled the model into Project...