Frontier LLMs match the favorite at 63.9%

Six frontier LLMs averaged 63.9% match-outcome accuracy across all 104 games in WorldCup Arena, merely matching the strategy of always backing the bookmaker’s favorite, according to the August 4 preprint. The prospective benchmark froze 4,494 predictions before kickoff, eliminating answer memorization; the models agreed more often than they were right, so majority voting added nothing. Builders should file that result beside evidence that harness choices can swing coding-agent rankings: leakage-free evaluation can expose apparent model diversity as correlated consensus rather than independent signal.