GPT-6 Astra Scored 99.9%—So Why Did OpenAI Lock It Away?

OpenAI says GPT-6 Astra is its most intelligent model ever. It also won't let you use the version that earned those numbers. On September 3, 2026, Astra posted results that looked like the end of the argument — as high as 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 100% on ExploitBench. Greg Brockman closed the press call with "Welcome to the AGI era." The full model is still restricted to vetted defenders. This video breaks down why both of those things are true at once: the 100,000-GPU Stargate training run in Texas, the multi-agent architecture that lets Astra run for hours, the $2,000 internal run that cracked 10 long-standing open math problems later verified in Lean, and the benchmark result that swings from 62% to 99% depending on nothing but the test harness. Then the part launch-day coverage moved past quickly: Astra is the first model OpenAI has ever classified "Critical" under its Preparedness Framework. It found two previously unknown zero-days during evaluation. The version shipping to Plus subscribers has those capabilities removed. The question isn't whether it can do your job. It's who gets the version that can break everything else. #OpenAI #GPT6 #Astra #AGI #AI