OpenAI's GPT-6 Astra model has secured the top position on the ErdosBench, a benchmark featuring 226 open mathematical problems. The model achieved a score of 3.23, successfully solving 106 problems, with 43 solved completely and 27 disproved.
The performance of Astra represents a 5% to 10% improvement in math-research capabilities compared to previous benchmarks. It outperformed the GPT-5.6 Sol model, which solved 78 problems, while demonstrating more accurate scientific writing and fewer exaggerated claims.
Despite these results, OpenAI chief scientist Jakub Pachocki stated that the company did not prioritize mathematical optimization for Astra. Resources were instead directed toward recursive self-improvement and automated alignment research to maintain a competitive edge in AI development.
This strategic shift highlights the trade-offs in AI development, where focusing on specific domains like math may lead to stagnation in other areas. The current trajectory suggests a move toward specialized models rather than broad, uniform improvements across all human tasks.
The rapid generation of proofs by AI models has prompted concerns among mathematicians regarding the future of the field. Experts warn of a potential shift from a scarcity of proofs to an overload, requiring a reevaluation of mathematical values and research practices.
