**Recent AI solutions to FrontierMath open problems, especially the August 12, 2026, Hadamard matrix of order 668 result credited to Anthropic’s Claude plus human collaborators, are the main catalyst shaping trader views.** Epoch AI, which maintains the benchmark of unpublished research-level math problems, has now recorded multiple AI-assisted solutions in 2026 after the first one in March using GPT-5.4 Pro. OpenAI’s August 1 release of ten advances via its internal Astra model further underscores accelerating large language model reasoning on frontier tasks. Progress on related benchmarks shows leading models (GPT-5.x variants, Claude Opus/Fable series, Gemini Deep Think) now solving 40–50+% of Tier 1–4 problems and approaching higher marks on cleaned sets, fueled by post-training on math domains, longer chain-of-thought, and hybrid human-AI workflows. Competitive dynamics among OpenAI, Anthropic, Google DeepMind, and others create incentives for rapid iteration, while the “wisdom of crowds” in prediction markets reflects skin-in-the-game assessment of these trends. Key upcoming catalysts include additional model releases or post-training updates through late 2026, potential expansions of the FrontierMath open-problem set, and the next round of the First Proof initiative. Resolution criteria hinge on whether Epoch AI credits core ideas to AI systems on new unsolved problems, with the market-implied odds heavily weighting continued momentum over historical patterns of benchmark saturation.