OpenAI's Astra Tops Math Benchmark Despite Skip

OpenAI's newest model, GPT-6 Astra, just posted the top score on ErdosBench, a set of 226 open math problems built by ulam.ai around the tradition of Erdős-style questions. Astra solved 106 of them, closed out 43 completely, and disproved 27 others, landing a benchmark score of 3.23. Benchmark creator Przemek Chojecki pegged the jump at roughly 5 to 10 percent across various math-research skills compared to the field's previous leader, Sol, which managed 78 solved problems at maximum reasoning settings. Astra also wrote up its results with less exaggeration, in some cases underselling what it had actually proven.

That result lands oddly given what OpenAI has said about its own priorities. The company built its first Astra announcement around math achievements, yet chief scientist Jakub Pachocki now says math research wasn't where the team focused its effort this time. Mathematicians who've spent the past stretch of time nervously watching model releases for signs their field was about to get automated away can exhale, at least for now. The gap between what OpenAI chose to build and what Astra turned out to be capable of raises its own questions about how the company is thinking about progress toward AGI.