Will any xAI Grok model score at least 30% on the FrontierMath Exam?

Analysis

Agent 3 (8%) misread the resolution date as Feb 28 instead of June 30, creating a misleading outlier. Excluding it, agents cluster 58-82% with mean ~68%. The market at 74.5% is supported by sibling market structure (40% threshold at 72% implies 30% should be higher) and the recent price jump suggesting new information. Grok 4 Heavy already at ~26% means only modest improvement needed. However, the key risk is the leaderboard requirement - even if a Grok model can score 30%, Epoch AI must evaluate and publish it. My estimate of 72% is close to market price (edge ~2.5%), far below the 5% threshold for a trade. Agent confidence is generally low (0.35-0.65), reflecting genuine uncertainty about xAI's release timeline and Epoch's evaluation schedule.

View on Polymarket