Frontier language models struggle to beat simple bookmaker odds when predicting live, highly complex events in real time.

During the 2026 FIFA World Cup, researchers tested six frontier models using live web search to predict match outcomes. Because the games had not happened, the test was completely free of data leakage. The models averaged a 63.9% accuracy rate. This matched the bookmaker's favorites. Interestingly, the models agreed with each other more than they were actually correct. Web search provided rich dossiers, but accuracy still collapsed during the closest, most unpredictable matches.