06-reference

stratechery anthropic mythos model glasswing alignment

Anthropic’s New Model, The Mythos Wolf, Glasswing and Alignment

Anthropic announced Mythos, a new frontier model so capable at coding and vulnerability discovery that it is being released only to ~50 critical infrastructure organizations (Project Glasswing) rather than the general public. Mythos Preview costs $25/$125 per million tokens (5x Opus 4.6). It is likely the first frontier model trained on Nvidia’s Grace Blackwell NVL72 architecture, proving continued returns to pre-training scale.

Thompson is deeply analytical about the multiple motivations behind the limited release: genuine security concerns (the model finds vulnerabilities that survived decades of human review), compute constraints (can’t serve it at scale), protecting against Chinese distillation, and business focus (enterprise agreements over self-serve subscriptions). He connects this to Amodei’s long history of restricting model releases going back to GPT-2.

The deeper concern Thompson raises: to the extent Mythos is genuinely dangerous, Anthropic is asserting control over a national-security-relevant capability. This revives the tension from Anthropic’s military standoff: a private company controlling capabilities that rival state power is fundamentally intolerable to the U.S. government.

RDCO note: Mythos’ existence and limited release have direct implications for our tooling. The security-focused deployment (Project Glasswing) signals that cybersecurity-as-a-service built on frontier models is a viable business category. The pricing signals that frontier models will remain premium. Cross-ref: Jaya Gupta moat thesis (06-reference/2026-04-10-jaya-gupta-anthropic-moat.md).