How the five AIs actually did

The Group Stage of the 2026 World Cup finished today, so it’s time to mark the homework. Back in the launch post five AI models, a deterministic lookup table, and one football fan predicted the same 72 fixtures before a ball was kicked. I’ve been scoring the results on AIWC26 each day as they came in. I’ve got no real opinion on the teams; I’m interested in what the scoring reveals about the models. Final table Entrant Points Exact scores Correct results Claude Sonnet 4.6 67 10 37 Claude Fable 5 64 9 37 Gemini 3.1 Pro 62 8 38 Gemini 3.5 Flash 62 10 32 Ranking Bot 59 7 38 GPT-5.5 52 6 34 Human 1.0 52 6 34 The lookup table-powered Ranking Bot beat one of the five models. Twenty lines of Python, no training data, no judgement. It still out-scored GPT-5.5 and the football fan who claimed to care about this in the first place. GPT-5.5 and the human didn’t just tie on points, they tied on the breakdown too, 6 exact scores and 34 correct results each, arrived at by a model reasoning…