The model that was too capable
In 2026 the most capable models find vulnerabilities better than almost anyone, and the three big American labs release them first, or only, to defenders.
Claude Mythos Preview and Project Glasswing (April 7, 2026)
Anthropic announces Claude Mythos Preview and decides not to release it. It gives the model to Project Glasswing so participants can fix their systems before models like it become widespread.
GPT-5.5 (April 23, 2026)
Meanwhile, OpenAI rates GPT-5.5 at the high level for cybersecurity in its Preparedness Framework and releases it with stricter safeguards.
Claude Fable 5 and Claude Mythos 5 (June 9, 2026)
Anthropic releases a Mythos-class model with safeguards, Fable 5, and reserves the version without them for Glasswing partners. Three days later, the U.S. government restricts access to it.
GPT-6 Astra (September 3, 2026)
GPT-6 Astra reaches the critical level for cybersecurity in OpenAI's framework and launches with limited access.
Gemini 4 Argon (September 30, 2026)
Google follows the same path. Gemini 4 Argon goes first to selected cybersecurity defenders.