AI model scores ≥ 90% on FrontierMath Benchmark before 2027?

In 2024, the best AI models in the world were solving single-digit percentages of FrontierMath problems — research-level mathematics that takes human experts hours or days. By mid-2026, the leading model sits near 50%. The market is pricing whether that trajectory continues at roughly the same pace for another six months. That's the entire bet. Not whether AI can eventually solve research-level mathematics — that debate is largely settled in the affirmative. Whether it gets there by December 31, 2026 specifically. FrontierMath was designed explicitly to resist AI — problems sourced from open research questions, constructed to prevent memorization and pattern-matching from training data. The fact that models have gone from near-zero to nearly 50% on a benchmark specifically engineered to stay hard is the most important context for understanding why the market prices YES so aggressively. Benchmarks designed to be AI-resistant have consistently fallen faster than their designers anticipated. The gap between current performance and the resolution threshold is real — roughly 40 percentage points separating the leading model from 90%. The question is whether that gap represents a genuine ceiling or a timing question. Expert forecasters at Epoch AI place 80-plus percent performance in the late 2020s, which would put 90% by end-2026 ahead of their median timeline. Markets are pricing faster than expert consensus — which has been the correct bet on AI capability development repeatedly across the past several years. The trajectory argument and the expert timeline argument pull in opposite directions, and the market has sided decisively with trajectory. That's not irrational given the track record of expert timelines in AI capability forecasting — they've consistently underestimated speed. Bottom line: This resolves on whether AI capability development continues its recent pace or hits a genuine inflection point in the second half of 2026. Watch FrontierMath leaderboard updates and any new frontier model releases as the direct leading indicators — a single major model release could move this contract dramatically in either direction.