OpenAI says GPT-6 Astra's math ability was a side effect, not the goal

GPT-6 Astra tops the ErdosBench for open math problems even though OpenAI's chief scientist Jakub Pachocki says math wasn't a development priority. OpenAI says it instead focused resources on recursive self-improvement and alignment research.

Why it matters: This supports a 'spiky' capability pattern where models leap ahead in domains that weren't even targeted, rather than improving uniformly across the board. It also signals OpenAI is explicitly betting on self-improving systems as its next frontier, a direction with far higher stakes for safety than benchmark chasing.