OpenAI’s October 6–7, 2026 release of 722 manuscripts from an unreleased internal frontier model—covering progress on the quasi-Riemann hypothesis, Unique Games Conjecture, and hundreds of other open problems—has sharply accelerated expectations for Epoch AI’s FrontierMath: Open Problems benchmark. Several results include Lean-formalized proofs, though independent verification remains ongoing and a handful of papers have already been withdrawn. This follows GPT-6 Astra’s September solutions to two Erdős problems and recent First Proof project successes on research-level questions, including a quantum topology problem solved by one of five tested systems. Competitive pressure from Claude Fable 5 and GPT-6 variants, which now post 88–100% on FrontierMath v2 Tier 4, underscores rapid scaling of large language model reasoning. Traders are watching Epoch AI’s next official classifications and any new Lean-verified submissions before year-end deadlines.