A new report titled 'KellyBench' from the London-based AI start-up General Reasoning reveals that leading AI models, including those from Google, OpenAI, and Anthropic, struggled with betting on soccer matches throughout a simulated 2023–24 Premier League season. The study tested eight AI systems, providing them with detailed historical data and statistics to build models aimed at maximizing returns while managing risk. Each AI agent was limited to three betting attempts and could not access the Internet for real-time updates. Anthropic’s Claude Opus 4.6 performed the best, with an average loss of 11% and nearly breaking even on one attempt. In contrast, xAI’s Grok 4.20 went bankrupt once and failed to profit in its other attempts. Google’s Gemini 3.1 Pro achieved a 34% profit in one attempt but also faced bankruptcy in another. These results underscore the challenges AI faces in adapting to dynamic, real-world environments over extended periods.