OpenAI's GPT-6 Astra Tops ErdosBench Despite Math Being 'Not a Priority,' Chief Scientist Says

OpenAI's GPT-6 Astra has taken the top spot on ErdosBench, a benchmark built around open problems in mathematics, according to reporting from The Decoder. The result would be a headline achievement for most labs. For OpenAI, it is something closer to a side effect: chief scientist Jakub Pachocki said math was deliberately not a focus during the model's development.

That admission reframes the benchmark win. OpenAI has instead been directing its compute and research staff toward recursive self-improvement and alignment work — the twin problems of getting a model to improve its own capabilities and keeping those capabilities pointed in a safe direction. Math, in this account, received no special optimization push. Astra cleared the benchmark anyway.

The gap between where OpenAI spent its effort and where Astra performed is the real story. It supports the view that AI development is becoming 'spiky,' with models posting extreme results in select domains while lagging elsewhere, rather than advancing evenly across the board. A model can beat mathematicians at their hardest open questions while still failing at tasks a human handles without thinking.

The mechanism behind that spikiness matters too. As long as systems cannot fully improve themselves, labs still need targeted optimization runs built on human-generated data to move a specific capability forward. That requirement limits how far any general model drifts from its training. It also means a strong benchmark score reflects a decision about where to point resources as much as it reflects raw model quality.

For mathematicians, the ErdosBench result is a mixed signal. If Astra can contribute to open problems without being trained for it, that suggests real utility for research workflows — and possibly more pressure on a field already sorting out how to credit and verify machine-assisted results. OpenAI has not said whether it plans to publish the specific problems Astra solved or the methods it used to reach them.

What happens next depends on whether OpenAI keeps math on the back burner. If recursive self-improvement remains the priority and spiky performance is a byproduct, expect more benchmark wins in domains the company never targeted — and more questions about which capabilities are genuinely robust versus which are artifacts of a narrow test. Pachocki's comments suggest OpenAI sees the unevenness as intended, not as a problem to fix.

The open question is whether that approach holds as models scale. A system that needs human-curated data to sharpen each domain will keep producing specialists, not generalists. Whether that counts as progress toward more capable AI, or a detour around the hard parts, is the argument OpenAI just handed to its critics.