Seven frontier AI models were given the 2026 World Cup draw and tasked with forecasting the champion. Four picked Spain (boldest: Stepfun at 33%), three picked defending champion Argentina (boldest: Qwen at 22%). All placed Spain, Argentina, and France in their top tier. The models used diverse methods: Opus 4.8 Max applied Dixon-Coles Poisson + Monte Carlo, factoring in heat and altitude; GPT 5.5 used weighted scorecards and cross-checked with Opta; DeepSeek v4 Pro wrote a 5,000-word epic but relied on outdated data; Stepfun 3.7 ran 50,000 Elo-based simulations, initially producing nonsense but correcting transparently. Key implications: AI prediction varies wildly based on methodology and data quality; no model is infallible, and biases (e.g., outdated rosters) can distort results. The exercise highlights both the power and limitations of AI in complex, human-centric domains like sports.

Key facts

• 4 of 7 AI models picked Spain; 3 picked Argentina as champion.
• Stepfun was most confident (33% for Spain); Qwen most confident for Argentina (22%).
• Opus 4.8 Max factored in heat, altitude, travel; cut Brazil to 8%.
• DeepSeek v4 Pro used outdated data (Southgate, Dorival) despite thoroughness.
• Stepfun's first model failed (Mexico top-3); rebuilt on pure Elo with transparent results.

KeyAudit data perspective