AI Models Are Betting on the World Cup in a New Prediction Benchmark

Obside is testing whether AI can predict uncertain outcomes by having major models bet on World Cup matches. Instead of using another standardized benchmark, the contest asks models to make decisions with incomplete information and live market odds.

The core question is simple and useful: can an AI turn public information into better predictions before anyone knows the result?

The models in the contest include ChatGPT, Gemini, Claude, Grok, Mistral, DeepSeek, and Kimi. An hour before kickoff, each model enters agent mode, researches teams, injuries, and other public information, then decides how much of a virtual $10,000 bankroll to wager using live Polymarket odds.

That setup makes the benchmark more like a judgment test than a trivia exam: the models must weigh evidence, uncertainty, and risk.

After the semifinals, Business Insider reported that Mistral led the field, followed by OpenAI’s GPT 5.5 and DeepSeek’s V4. Claude Opus 4.8 was at the bottom and was the only model in the red.

The rankings matter less as a final verdict than as a snapshot of how different AI systems handle live, ambiguous decisions.

The contest highlights a weakness in many AI evaluations: they often test knowledge or task completion, not judgment under uncertainty. Betting on soccer is not the same as broader forecasting, but it creates a measurable way to see whether models can make profitable predictions from online information.

Business Insider also notes that ChatGPT previously entered a secret forecasting tournament run by economists and performed no better than the average human contestant.