**Rapid gains in frontier AI reasoning systems have positioned models close to the 90% threshold on FrontierMath, driving the strong 90% market-implied odds for a breakthrough before 2027.** FrontierMath, developed by Epoch AI, consists of hundreds of original, unpublished research-level mathematics problems across tiers of increasing difficulty, designed to resist contamination and test expert-level capabilities that previously required hours or days of human mathematician effort. As of mid-2026, top models have advanced dramatically: Claude Fable 5 (max) reached 87.8% on the updated Tier 4 (v2) benchmark released in June 2026, while OpenAI’s GPT-5.6 Sol variants have posted scores in the low-to-mid 80s on the same set and up to 89% on legacy snapshots. These results reflect iterative improvements in long-horizon reasoning, chain-of-thought techniques, and domain-specific training for mathematical domains such as number theory and algebraic geometry. Traders see continued momentum from the ongoing release cycle of advanced large language models (LLMs) and reasoning systems through late 2026, including potential GPT-6 iterations or refined Claude and Gemini releases. Historical precedent shows scores rising from under 2% at the benchmark’s 2024 launch to over 40% on easier tiers and near 90% on harder ones within roughly 18–20 months. Key swing factors include further scaling of inference-time compute, algorithmic refinements, and possible specialized “AI co-mathematician” training runs. While timelines can slip and exact benchmark versions matter for resolution, the current trajectory and proximity to the threshold underpin the high consensus that 90% will be achieved before the end of 2026.