Google AI Co-Mathematician Sets FrontierMath Tier 4 SOTA

谷歌「AI联合数学家」来了！刷新最难数学AI基准SOTA，牛津教授用它解开群论悬案

Google DeepMind released AI Co-Mathematician, an asynchronous agent workspace for math research, and answered 23 of 48 private FrontierMath Tier 4 problems, scoring 48% under 48-hour, no-token-limit conditions versus GPT-5.5 Pro at 39.6%.

Why it matters: HKR-H/K/R all pass: the story has a hard benchmark number and a concrete research hook. No disclosed product access or cross-source cluster, so it stays at the top of 78–84 rather than p1.