Anthropic's Most Powerful Model Is Too Dangerous to Release Publicly

Anthropic unveiled Claude Mythos Preview — its most capable model yet — and immediately decided the public can't have it. Instead, they built Project Glasswing: a controlled rollout to AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, JPMorgan, Palo Alto, and 40+ other orgs to harden software before bad actors get anywhere near the model.

Mythos hits 93.9% on SWE-bench (vs. Opus 4.6's 80.8%) and has already surfaced thousands of high-severity vulnerabilities — including a 27-year-old bug in OpenBSD. Anthropic is backing it with $100M in usage credits to cover the remediation work across partner organizations.

The contrast with OpenAI is pointed: a day after OpenAI published a 13-page policy essay on cyber threats, Anthropic shipped working fixes. Project Glasswing is explicitly a learn-before-you-scale experiment — Anthropic says general availability isn't the goal yet.

Why it matters: the era of 'release and patch later' may be ending — frontier labs are now gatekeeping their most capable models and deploying them as infrastructure-hardening tools before public launch.