Can AI Beat the Sports Betting Market? 8 of the Top Models Tried

KellyBench tested eight leading AI models including Claude, GPT-5, Gemini, and Grok on Premier League sports betting predictions over a full season, and none generated profits. The results highlight the persistent difficulty AI faces in beating efficient markets despite advances in language models and reasoning capabilities.

The KellyBench experiment represents a rigorous stress test of AI's real-world predictive capabilities under monetary incentive. Despite rapid improvements in large language models, the failure to achieve profitable betting performance across eight distinct models reveals fundamental limitations in applying AI to dynamic, adversarial environments like sports markets. These markets embed sophisticated human judgment, incomplete information, and constant recalibration by professional bettors, creating conditions where raw pattern recognition struggles.

This result challenges the narrative that scaling AI capabilities automatically translates to economic advantage. Sports betting markets have historically been targets for machine learning applications because outcomes are quantifiable and historical data abundant. Yet the consistent zero-profit outcome across models of varying architectures suggests that language-based reasoning alone cannot overcome the inherent unpredictability of sporting events or the efficiency of modern bookmaking algorithms.

The broader implication extends to cryptocurrency and financial markets, where similar AI applications are being deployed. If leading AI models cannot consistently extract edge from sports betting—a domain with clearer rules and less manipulation than crypto markets—skepticism is warranted regarding AI-driven trading systems' actual performance versus marketing claims. This compounds concerns about AI agents managing capital in volatile, unregulated crypto ecosystems.

The experiment underscores that market efficiency and adversarial adaptation remain formidable obstacles. Future AI applications in prediction should focus on asymmetric information opportunities rather than competing in efficient public markets where computational sophistication alone provides minimal advantage.

• →Eight AI models including Claude and GPT-5 failed to achieve profitable returns over a full Premier League betting season
• →Language models struggle to beat markets where professional participants already exploit available information efficiently
• →AI's failure in sports betting raises questions about real-world performance of AI-driven trading systems in crypto and traditional finance
• →Market efficiency and adversarial adaptation remain fundamental barriers that scale and architecture alone cannot overcome
• →Profitable AI applications likely require asymmetric information or structural market inefficiencies rather than public data analysis