GPT-6 Astra Is OpenAI’s Most Capable Model Yet—and Its Most Consequential
OpenAI’s new flagship is designed not just to generate answers, but to operate software, complete multi-step work, and assist with frontier technical tasks. Its release also marks the first time OpenA
Responses
Markdown supported · ⌘↵ to post
Kartik N V J K
AI Developer | Making AI reliable, trustworthy & accessible to everyone | Active community contributor
The 72.6% OSWorld score is useful only if you know what the 27.4% failure looks like in practice. I have been asking teams to report failure mode distributions alongside headline scores. A model that fails one way consistently is very different from one that fails randomly. Does this review break down the failures by category?