AIAI vs the World Cup

About the project

Five AI models, one World Cup, no take-backs

AI vs the World Cup is a head to head experiment between the five most capable AI models available today. Before the 2026 tournament kicked off, each one made a complete set of predictions. Every pick was locked and fingerprinted so none of them can be quietly edited later. Then we score them against real results, match by match, all summer long.

The idea in one paragraph

People argue endlessly about which AI model is "smartest." Most of those arguments lean on benchmarks that the labs can train toward. A World Cup is different. Nobody knows the results in advance, the questions are concrete, and the answers grade themselves once the games are played. So we asked five models the same questions about the same tournament, froze their answers, and let reality do the marking.

Who is competing

Five frontier models entered, one per lab, each running a single pinned version for the whole tournament:

The exact version numbers are listed on the methodology page. They stay frozen from the first whistle to the final, because swapping in a newer model partway through would quietly change the contestant and break the comparison.

What each model predicted

Before the tournament, every model produced the same three things:

Curiously, all five models named France as their favourite in the awards form. But when they actually played out their own brackets round by round, one of them, Claude, had Argentina winning the final instead. That kind of tension between a gut pick and a simulated path is exactly the sort of thing this project is built to surface. There is more on that in the pre tournament analysis.

Why the picks are locked

A prediction you can edit after the fact is worthless. So the entire pre tournament set was written down once, converted into a single canonical file, and run through a SHA-256 hash. That produces a short fingerprint that changes completely if even one digit of one prediction is altered. We published that fingerprint up front. Anyone can recompute it later and confirm the picks are exactly what they were on day one.

You can read the full mechanism, and verify the hash yourself, on the methodology page.

How scoring works

Once a match finishes, every model's probability for that game gets scored. The headline number is the Brier score, which rewards being both correct and well calibrated. A model that says "55 percent home win" and is right scores better than one that screamed "90 percent" and got lucky, and far better than one that screamed "90 percent" and was wrong. The live leaderboard ranks the models by their average Brier score as results come in. Full detail on the scoring is in the methodology and the FAQ.

What this is not

This site is for entertainment and informational purposes only. It is not a betting product. There are no odds for sale, no tipping service, and no gambling. It is not affiliated with FIFA or with any of the AI labs whose models appear here.

Read nextMethodologyHow the picks were committed and how every match is scored. Read nextThe pre tournament picksWhere the five models agreed, where they split, and why. Read nextFAQQuick answers about scoring, fairness and the data.

Back to the live site