Locked in: our pre-tournament World Cup predictions

The World Cup kicks off today at the Azteca. Before the opening whistle, we did the thing this whole project has been building toward: we locked in every prediction. All 72 group-stage fixtures, five models each — 360 individual predictions — logged and timestamped this morning, hours before kickoff. No edits, no hindsight. Whatever happens over the next five weeks, the record was written first.

Who wins the whole thing?

We ran 10,000 tournament simulations per model on this morning's refreshed predictions. The five-model average makes England the narrowest of favourites, with Argentina close enough that the two are effectively co-favourites — and nobody clearing 10%. A 48-team World Cup is a genuinely open one.

Five models, four different answers

The average hides the more interesting story: the models don't agree on who lifts the trophy.

• Claude is the biggest England believer in the field (11.5%), and GPT-5.5 agrees with the pick.
• GPT-5.4 goes the other way: Argentina at 11.1%, its clear number one.
• Grok refuses to choose — England and Argentina sit in a dead heat at 10.0% each.
• Gemini is the contrarian: its favourite is Portugal at 8.2%. It rates England at just 7.0% and Argentina at 5.0% — barely half what the GPT models give them.

That Argentina number is the widest disagreement on the board: a 6.1-point spread between GPT-5.4's 11.1% and Gemini's 5.0%. Portugal is the opposite — the tightest consensus in the top ten (1.1-point spread), every model placing them between 7.4% and 8.4%. If you're looking for the field's most confident collective opinion, it's not the favourite. It's Portugal, third.

And then there's Brazil: ninth, at 3.9%, with remarkably little disagreement. Five models trained on different data by different labs all looked at the five-time champions and shrugged. That's either the most damning consensus of the pre-tournament board or the one it will most enjoy disproving.

Tonight's opener

Even here the fingerprints show. Grok gives Mexico 66% — ten points above the most cautious model — which is consistent with what our calibration study found: Grok leans hardest on market odds, and the market loves a host playing at the Azteca. Claude's 58% is notable for the opposite reason: the model our research flagged for over-pricing home advantage is not the high outlier tonight. The first of many small tests.

How predictions run during the tournament

From today, the system runs on a fixed cadence, all timestamped before play:

• Daily, early morning: a fresh 10,000-simulation tournament snapshot per model — the tournament page always shows the latest.
• Daily, hours before the first kickoff: match-level predictions from all five models for every upcoming fixture, on the matches page.
• Weekly: a deep refresh where every model re-forms its opinion of every possible matchup, absorbing form, injuries, and results.

After each match, predictions get scored against the result. The leaderboard will tell us, in public and in real time, which model actually understands tournament football — and whether knowing their biases beats trusting their averages.

The honest caveats

These are pre-tournament numbers built on raw model probabilities — group opponents only, knockout paths simulated over every possible pairing. The 32 knockout fixtures will be predicted and logged as their matchups become real. And as we wrote in our limitations post, friendlies and league calibration are not World Cup evidence. Starting tonight, we get the real thing.