GPT‑5.2 Pro cracks Erdős problem in landmark AI math first

This Erdős #281 result is one of the clearest examples yet of a frontier model doing genuinely new, checkable mathematics rather than just tidying up known arguments. GPT‑5.2 Pro, used by Eclipse founder Neel Somani, generated a proof that top mathematicians, including Terence Tao, judged correct for a problem that had been open since 1980.([tech.sina.cn](https://tech.sina.cn/2026-01-18/detail-inhhtasx8868326.d.html?oid=WA+0812+2782+5310+Biaya+Renovasi+Plafon+Model+Sekarang+Murah+Jebres+Solo&vt=4)) OpenAI has already been positioning GPT‑5.2 as a tool for advanced science and math, and this case study concretely backs that narrative by showing the model navigating subtle measure-theoretic and ergodic arguments without obvious hallucinated gaps.([openai.com](https://openai.com/index/gpt-5-2-for-science-and-math/?utm_source=openai))

For the race to AGI, this is a psychological and strategic inflection point. Industrial labs are no longer just benchmarking models on synthetic math tests; they are now touting solutions to long‑standing open problems as proof of capability. That will push competitors to demonstrate similar feats in other fields—formal methods, algorithm design, even parts of theoretical physics—driving an arms race around “AI as co‑author of frontier research.” At the same time, Tao’s own caution, echoed in Chinese and Indian coverage, that many solved problems may be from the “long tail” of underexplored questions is a reminder that this is still tightly human‑supervised collaboration, not autonomous discovery.([news.softunis.com](https://news.softunis.com/50647.html)) The labs that win this phase will be those that can institutionalize workflows where models explore and humans verify at scale.