OpenAI released GPT-6 Astra on September 10, 2026, saying it tops the ErdosBench for open math problems, even though chief scientist Jakub Pachocki says the company didn't prioritize math. It is the company's first major math-focused update since its initial announcement.
OpenAI reported a score of 3.23 on ErdosBench, measured on 226 open math problems inspired by the famous Erdős problems. That compares with Sol's score of 3.25, which has since retaken the top spot with better overall performance.
GPT-6 Astra is built on OpenAI's internal research and targets math research, with availability beginning for researchers and mathematicians. The model solved the most problems overall, though it showed stronger scientific writing and was less prone to overblown claims.
"We believe we could make the models better at specifically mathematics research with additional focus, but we do not prioritize this direction because of the urgency we feel about RSI and automated alignment research," said Jakub Pachocki, chief scientist. He added that OpenAI is pouring its resources into recursive self-improvement and securing future AI systems.
The announcement follows OpenAI's focus on recursive self-improvement and AI safety, as the company frames its priorities. The release highlights the trade-offs in AI development, where optimization in one area may mean less progress in others.
OpenAI did not say it will prioritize math optimization in the future, and mathematicians remain uncertain about AI's potential for real breakthroughs. The field is still set to evolve, even if some researchers doubt AI's ability to deliver genuine self-improvement.
Source: thedecoder
