OpenAI’s latest model, GPT-6 Astra, has taken the top spot on ulam.ai’s ErdosBench, solving 106 open mathematical problems and completely resolving 43 of them. It also disproved 27 others, earning a score of 3.23. This performance marks a solid 5%-10% gain in research skills compared to previous iterations, according to benchmark developer Przemek Chojecki. Yet, this dominance comes with a caveat that should give pause to anyone tracking the race toward Artificial General Intelligence. OpenAI didn’t actually try to win at math.
Chief scientist Jakub Pachocki admitted in his essay “An Alien Mind” that the company deliberately chose not to optimize Astra for mathematical research. Instead, resources were diverted toward recursive self-improvement (RSI) and automated alignment research. Pachocki argues that securing future AI systems is the only way to remain at the frontier. This strategic pivot suggests that today’s most capable math model is merely a byproduct of broader architectural goals. It implies that specialized excellence no longer requires specialized tuning, but rather emerges from general reasoning capabilities focused on self-correction.
This development lends credence to the “spiky” AGI thesis proposed by researchers like Adam Hunt. Rather than a smooth, gradual improvement across all human tasks, AI progress appears jagged. Models may achieve superhuman performance in coding or mathematics while stagnating in social reasoning or common sense. OpenAI’s trade-off highlights a hard reality: even leading labs cannot push maximum progress in every direction simultaneously. Capabilities grow where you point the optimization lens, and pointing it elsewhere means accepting gaps in other domains.
The implications for mathematicians are mixed. While Astra provides a powerful tool, it exacerbates a growing crisis of verification. As Terence Tao noted at the 2026 International Congress of Mathematicians, the field risks shifting from proof scarcity to proof overload. If AI generates results faster than humans can validate them, the core value of mathematics shifts from discovery to curation. We are entering an era where the bottleneck isn’t solving the problem, but deciding if the solution actually matters. That is a philosophical hurdle no amount of compute can clear.
