OpenAI's GPT-6 Astra leads ErdosBench math benchmark, despite no optimization
openai's gpt-6 astra tops erdosbench with 106 solved math problems, even though the company prioritized recursive self-improvement over math research.
openai's gpt-6 astra model has taken first place on the erdosbench benchmark for open math problems, scoring 3.23 and solving 106 problems, 43 of them completely. it disproved 27 more problems and showed stronger scientific writing compared to gpt-5.6 sol. however, openai did not prioritize math research for astra, instead focusing on recursive self-improvement and automated alignment research. chief scientist jakub pachocki stated that the company believes it could make models better at math research with additional focus, but chose not to. this decision reflects the trade-offs being made at the frontier of ai development, where optimizing for one area may mean sacrificing capabilities in another.