GPT-6 Astra 在 ARC-AGI-3 上跑出 99.9%

Token Gremlin @TokenGremlin原推发布于 2026.09.04

在 X 上打开原推

One thing I can’t stop thinking about: ARC-AGI-3 was designed as a benchmark where humans could solve the environments, while frontier AI initially scored around 0.5%. GPT-6 Astra just hit 99.9% with OpenAI’s Provider Adapter. Yes, the harness matters. The standard score is lower. But watching a benchmark go from “AI can barely do this” to effectively saturated this fast is honestly pretty unsettling. Greg Brockman is calling Astra the beginning of the AGI era. If the real-world model is actually this good, I’m starting to understand why.