AI models are terrible at betting on soccer—especially xAI Grok

A recent study has revealed that even the most sophisticated AI models from tech giants like Google, OpenAI, and Anthropic faced significant challenges when attempting to predict outcomes in soccer betting over a full Premier League season. The research, conducted by the AI startup General Reasoning, emphasizes the limitations of AI in applying its advanced analytical capabilities to real-world scenarios over extended periods. The report, known as the "KellyBench," evaluated eight leading AI systems in a simulated environment of the 2023-24 Premier League season. These AI agents were provided with extensive historical data and statistics related to teams and past matches. Their objective was to create models aimed at maximizing profits while effectively managing risk during betting. During the experiment, the AI systems placed bets on match outcomes and the number of goals scored, testing their adaptability to evolving events and player data throughout the season. Notably, the AI models operated without internet access to retrieve real-time results, and each system was allowed three betting attempts. Among the participants, Anthropic’s Claude Opus 4.6 emerged as the most successful, recording an average loss of just 11 percent and nearly breaking even on one of its attempts. Conversely, xAI’s Grok 4.20 faced significant setbacks, going bankrupt during one attempt and failing to complete the other two. Google’s Gemini 3.1 Pro managed to achieve a 34 percent profit on one occasion but suffered bankruptcy in another attempt, highlighting the unpredictable nature of sports betting even for advanced AI systems.