OpenAI’s new model aces the benchmarks and admits it is better at hiding

OpenAI released GPT-6 Astra, its most advanced and aligned model, achieving a 99.9% score on ARC-AGI-3 and large gains across benchmarks such as 97.6% on FrontierMath Tier 4 and 57.9% on Terminal-Bench 4.0, while completing OSWorld 2.0 tasks 47% faster. Astra also hit 100% on ExploitBench, discovered two new zero-day flaws, and reduced hallucinations to 4.2%, though its reasoning can more easily evade oversight, which OpenAI flags as an open research priority.