Recent advances in frontier large language models have lifted top FrontierMath scores from roughly 25% in late 2024 to around 52% for GPT-5.5 Pro and 47–50% for GPT-5.4 variants as of May 2026, reflecting stronger chain-of-thought reasoning on research-level math problems. Yet this steady incremental progress—driven by scaling and post-training refinements—leaves a substantial gap to the 90% threshold. Traders assign 77.5% probability to “No” before 2027 because current trajectories and historical benchmark saturation patterns indicate that closing the remaining distance would require unprecedented reasoning leaps unlikely to materialize in the next seven months, even with anticipated releases such as GPT-6 or Claude 5. Upcoming model evaluations and any sudden capability jumps remain the main swing factors.