OpenAI’s GPT-6 Astra recently drove sharp sentiment shifts by posting 97.6–98% accuracy on FrontierMath Tier 4, prompting Epoch AI to label the result a “Major Advance” in September 2026—the first such designation. The model also solved two of 68 historically open Erdős problems shortly after that benchmark launched on September 1. A separate Apéry-style irrationality problem was marked solved later in the month under Epoch’s new “human-plus-AI” category, reflecting iterative collaboration rather than fully autonomous proof. These milestones follow June 2026 corrections to the core benchmark and continue a summer wave of AI-assisted research results. Traders are watching OpenAI’s internal agent systems and upcoming evaluations on the harder Erdős and open-problem sets for the next resolution triggers.