智能体工作是 Step 5 Preview 落后于同等 Intelligence Index 分数模型的地方。在 GDPval-AA 上它落后于 GLM-5.3（max）和 Qwen3.8 Max（1,566 Elo 对 1,646 和 1,668），在 AA-Briefcase 上（1,432 对 1,525 和 1,640），在 Terminal-Bench 4.0 上（33% 对 42% 和 39%）。在知识和推理方面则相反，它在 Humanity's Last Exam（46% 对 42% 和 43%）、CritPt（21% 对 19% 和 18%）和 AA-Omniscience Index（16 对 14 和 12）上均领先两者。
Agentic work is where Step 5 Preview lags models at a similar Intelligence Index score. It sits behind GLM-5.3 (max) and Qwen3.8 Max on GDPval-AA (1,566 Elo against 1,646 and 1,668), AA-Briefcase (1,432 against 1,525 and 1,640) and Terminal-Bench 4.0 (33% against 42% and 39%). The reverse holds on knowledge and reasoning, where it is ahead of both on Humanity's Last Exam (46% vs 42% and 43%), CritPt (21% vs 19% and 18%) and the AA-Omniscience Index (16 vs 14 and 12)
来源：@ArtificialAnlys · x.com