OpenAI's GPT-6 Astra Tops Math Benchmark Despite Deliberately Skipping Math Optimization

OpenAI's GPT-6 Astra has taken the top spot on ErdosBench, a benchmark of 226 open math problems modeled on the famous Erdős problems, scoring 3.23 and solving 106 problems outright — 43 of them completely. The model also disproved 27 additional problems.

That performance beat GPT-5.6 Sol, which scored 3.12 and solved 78 problems at maximum reasoning settings on the same benchmark, according to ulam.ai. Benchmark developer Przemek Chojecki characterized the jump as "a solid 5%-10% gain on various math-research skills tested," while noting the benchmark remains far from saturated.

What makes the result notable isn't just the score — it's that OpenAI says it didn't try to maximize it.

OpenAI chose not to optimize for math

In an essay titled "An Alien Mind," OpenAI chief scientist Jakub Pachocki wrote that the company "could make the models better at specifically mathematics research with additional focus," but chose not to prioritize that direction. Instead, according to Pachocki, OpenAI is concentrating resources on recursive self-improvement (RSI) and automated alignment research, which he describes as "the only way to remain at the frontier of AI research moving forward."

That statement means GPT-6 Astra's math performance is a byproduct of broader capability gains rather than a targeted push. It also confirms something researchers have suspected: even OpenAI faces hard trade-offs at the frontier. Pushing one capability area — in this case, math — means pulling resources from elsewhere.

Compared to Sol, Astra reportedly showed stronger scientific writing quality and was less prone to overstating its findings, in some cases understating its own results, according to the source reporting on the benchmark.

A spiky path to capability, not a broad one

The result feeds into a broader debate about how AI capability is actually developing. Cambridge researcher Adam Hunt has visualized two competing trajectories: a "mainstream AGI" thesis where models improve broadly and gradually across all human tasks, versus a "spiky" trajectory where models achieve extreme strength in narrow domains like coding and math while stagnating or regressing in areas like language quality, common sense, or social reasoning.

Pachocki's admission that Astra's math strength is incidental — not the product of deliberate optimization — is being read as evidence for the spiky thesis. If even the leading lab can't advance every capability simultaneously, general intelligence in the broadest sense remains further off than raw benchmark scores suggest.

Mathematicians face a different problem: too many proofs

At the 2026 International Congress of Mathematicians, mathematician Terence Tao raised a separate concern: as AI models generate proofs faster than humans can verify them, the field risks shifting from a scarcity of solutions to an overload of unverified results. The bottleneck, in Tao's view, could shift from solving problems to deciding which results are actually significant — a shift he compares to the foundational crisis mathematics faced in the early 20th century.

The hardest open problems in mathematics remain unsolved, buying the field time. But skepticism persists among mathematicians who doubt language models can produce genuinely creative breakthroughs rather than incremental extensions of existing techniques.

What this means

GPT-6 Astra's ErdosBench win demonstrates that even unoptimized models from frontier labs now outperform specialized math tools on hard research problems — a sign of broad capability gains rather than narrow tuning. But OpenAI's own account undercuts any narrative of steady, uniform progress toward general intelligence: resources are being explicitly redirected toward recursive self-improvement and alignment work, not mathematics. That's a bet that automating AI research itself will eventually yield gains everywhere, including math, faster than direct optimization would. Whether that bet pays off — and whether verification, not generation, becomes math's real bottleneck — remains an open question the field is only beginning to confront.