Rapid progress on FrontierMath, a benchmark of expert-level mathematics problems, drives the 99.7% market-implied probability that an AI model reaches ≥90% before 2027. As of early September 2026, OpenAI’s GPT-5.6 Sol leads the legacy version at 89% and FrontierMath v2 Tier 4 at 83%, following consistent gains from sub-2% at launch through GPT-5.5 releases and a June 2026 v2 update that corrected errors in 42% of problems and lifted scores across the board. Large language models continue to close the gap through scaled reasoning, tool use, and iterative releases, with trader consensus reflecting the historical pattern of math benchmarks saturating quickly once frontier systems exceed 50-70%. Remaining uncertainty centers on whether a new harder tier or version could reset the threshold before year-end.