Which AI is best at football predictions?

Any page that names one permanent winner is describing a snapshot; the real ranking changes week to week. Football results carry so much irreducible randomness that over any short window the difference between frontier models is a handful of matches — well inside the noise. That is why ScoreGPT publishes the comparison as a live leaderboard instead of a verdict: every model's pick is locked before kickoff, graded against the real result after full-time, and displayed over 7, 14 and 30-day windows. Whoever is on top today earned it on graded picks you can inspect — and may not be on top next week.

Same matches, same information, same grading — the comparison only means something because the conditions are identical. Each model receives a full pre-match dossier (form, injuries, tactics, stakes) plus live web search, runs at high reasoning settings, and files its pick before kickoff.

No model is dropped for a bad run and none is promoted for a good one — the roster changes only when a lab ships a stronger successor.

At the 2026 World Cup, four of the five sealed pre-tournament brackets named the exact final — Spain v Argentina — five weeks before it was played. Across more than 100 graded tournament fixtures, match-result accuracy ran between roughly 60% and 70% depending on the model, and Grok — then version 4.3 — posted the best graded record of the five. That is a dated, tournament-specific result, preserved in full on the World Cup 2026 AI report card — not a permanent ranking. It shows what a properly resourced panel of frontier models can do on a defined slate; the club season is a longer, harder grind.

The more useful question is: where do strong models agree, and where do they split? When five independently run frontier models read the same match and land on the same result, that consensus carries real weight — and when they split, the disagreement itself tells you the match is genuinely open. That is a signal no single "best" model can give you, and it is how ScoreGPT turns five opinions into one clear call. If you want one model to follow personally, read each model's behaviour on its record page and pick the analyst whose style you trust — then check it against the consensus.

Four checks settle it: (1) Are all picks published before kickoff, losses included? (2) Is the market named — three-way results are far harder than two-outcome markets? (3) Is the window dated, with the sample size shown? (4) Could you recompute the ranking yourself from the listed picks? ScoreGPT's record passes those checks by construction — every pick locks pre-kickoff and grades in public — and the same standard applies to anyone else claiming a winner. The companion page How accurate are AI football predictions? covers the baselines that make any percentage meaningful.