Study finds top AI models lost money betting on Premier League games

In short: A new test found that several leading AI systems lost money when asked to bet on Premier League matches across a full season.

What happened

A London-based start-up called General Reasoning released a report called KellyBench. It tested eight well known AI models from companies including Google, OpenAI, Anthropic, and xAI.

The company ran a virtual replay of the 2023-24 Premier League season. The AI systems got detailed historical data and team statistics, then they had to place bets on match results and the number of goals.

Each AI started with a simulated £100,000 bankroll, which is like a betting wallet. The AIs could not use the internet to look up real results as the season played out, and each got three tries to make a profit.

General Reasoning said every model it tested lost money on average. Anthropic’s Claude Opus 4.6 did best, with an average loss of 11%. OpenAI’s GPT-5.4 lost 13.6% on average. xAI’s Grok 4.20 lost everything, and Google’s Gemini 3.1 Pro had mixed results, including one run with a 33.7% profit and another where it went to zero.

The report authors said the AI systems “systematically underperform” humans in this setup. The paper has not yet been peer reviewed, which means other independent experts have not formally checked the methods.

Why it matters

People often see AI doing well on short tasks, like writing a paragraph or code. This study suggests that when the job requires making many decisions over time, with new information and surprises, today’s AI can struggle, a bit like a student who does fine on a quiz but falls behind in a long course.

Source: Financial Times