GPT-6 Astra 在 Agent's Last Exam 登顶
伯克利新基准用 1500+ 真实行业任务考 AI，Astra 击败 Anthropic 旗舰且更省 token。
There's now a benchmark that doesn't ask an AI trivia questions. It asks it to do your job. UC Berkeley and the RDI Foundation built "Agent's Last Exam": 1,500+ tasks, curated by 300+ industry experts, across 55 sub-industries, each graded on a verifiable, real-world outcome. Not multiple choice. Real software, real tolerances, real consequences. One task: open real industrial CAM software, turn a metal block into a finished part. 10,000 surface checks, a 0.3mm tolerance, zero crashes allowed. One collision and the score is zero, no matter how good the rest of the plan looked. GPT-6 Astra just set a new state-of-the-art on this exam, beating Anthropic's frontier model, and used fewer tokens to do it. If you run ops at a neobank or a crypto-card product, this is the part of the Astra story that should actually matter to you. Not "is it AGI." Whether the model can sit in your reconciliation spreadsheet, your compliance queue, your engineering backlog, and do graded, verifiable, economically valuable work without you watching over its shoulder. The benchmark is a preview of the job description.