Fields Medalist says ChatGPT 5.5 Pro delivered "PhD-level" math research in under two hours with zero human help

The detail worth sitting with is not the result itself but the confirmation source: MIT researchers independently verified the insight as novel, which is a meaningfully higher bar than a researcher saying 'this looks right.' That external validation step is what separates a compelling demo from a documented capability claim.

This lands differently when read alongside the MIT study from early May explaining why scaling language models works so reliably, specifically the finding that superposition drives predictable capability gains. Mathematical reasoning at this level was a predicted destination on that scaling curve, not a surprise detour. More broadly, the Harvard diagnostic accuracy story from the same week established a pattern: peer-reviewed or expert-validated AI performance claims are arriving faster than the institutions affected by them can update their norms. In mathematics, the downstream pressure is on journals and PhD programs to redefine what constitutes original contribution, a slower-moving institutional problem than regulatory approval timelines in medicine but no less consequential.

Watch whether a major mathematics journal, specifically one where Gowers has editorial standing, issues formal guidance on AI-assisted proofs within the next six months. If that happens before OpenAI publishes a technical report on the underlying reasoning architecture, it signals that institutional response is outpacing transparency from the model developer.