GPT-5.5 solved doctoral-level math without human help
The Path to AGI: Each Company's Approaches and Current Position2026年5月、フィールズ賞受賞者ティモシー・ゴワーズが「GPT-5.5 Proが人の助けなしで2時間以内に数論の博士レベル研究を完了した」と証言（指数関数的上界を多項式的上界に改善、着想は「完全に独創的」とMIT研究者が評価）。数学研究への参入はAGIへの道の重要な指標とされており、この到達点は「優秀な相棒」から「研究パートナー」への移行を示唆する事例として注目されている。
Six months ago, GPT-class models hit a ceiling on genuinely novel math — producing plausible steps, but nothing a Fields Medal winner would confirm as correct. GPT-5 launched three months ago and raised accuracy on olympiad-level problems, but inventing new theorems at doctoral level was still widely considered out of reach. Today, a Fields Medal winner confirmed that wall has come down.
For researchers in math and engineering, AI as an active research partner just became credible — not just a search engine, but a collaborator that can help draft proofs and surface novel ideas. On the business side, this is a landmark case for AI taking on parts of research roles, which will start influencing R&D investment decisions and hiring plans. For most people using AI in everyday work, today's GPT-5.5 isn't suddenly smarter — this is about understanding the ceiling of what's possible, not a daily-use upgrade.