Recently, an interesting experiment was conducted—multiple large models were allocated $10,000 each to trade over 6 weeks in a football prediction market. The results were quite dramatic.

GPT-5.1 led the pack with a 42.6% increase, followed closely by DeepSeek with a 10.7% profit, and Gemini 3 Pro remained steady at 5.5%. Opus 4.2 contributed 3.9%, while Grok 4.1 Fast achieved 2.1%. However, GPT-5.2 faltered, dropping by 21.8%—it seems not all models excel in this area.

This comparative test was jointly promoted by a prediction market platform and an AI research team. The underlying logic is quite interesting: testing the performance of different AIs in non-standardized decision-making tasks using real funds. Football prediction markets involve data analysis, probability estimation, and risk judgment—making it an ideal scenario to evaluate the practical trading capabilities of large models. The significant differences also reflect that having parameters and training scale alone does not guarantee market decision-making ability; execution strategies and data understanding quality are equally critical.