England 4–2 Croatia: Every AI Called the Winner — and Missed the Goals

Same brief, same scoreboard, public grading. This time the panel got the headline right — and the whole story wrong.

Two nights, two very different lessons. A day after a near-unanimous AI panel was held to a draw by DR Congo with Portugal, the same lineup lined up behind England against Croatia — and this time the favourite delivered. England won 4–2. Every one of the thirteen models that called the winner got it right. And yet, dig one market deeper, and the panel misread the game almost completely. They did not just expect England to win; they expected England to win boringly. Croatia, and a 4–2 goal-fest, had other ideas.

Here is the full, graded story of England vs Croatia — the calls the AIs nailed, the ones they whiffed, and what the split tells you about where these models are sharp and where they are soft.

TL;DR

• All 13 models picked England to win, at 53–65% confidence — and all 13 were correct. England won 4–2.
• But the panel expected a tight, low-event game: Under 2.5 goals (10 of 13), both-teams-to-score "No" (12 of 13), and a 1–0 scoreline (10 of 13).
• It was a 4–2 thriller. Every Under, every "No", and every 1–0 lost.
• Net of 91 graded predictions across seven markets: 22 won, 45 lost, 24 void. Right team, wrong game.

The matchup

England came in as favourites against a Croatia side that has built a tournament reputation on grinding out tight, low-scoring games. That reputation clearly shaped the panel's thinking — the models liked England to win, but they priced the match as a cagey, controlled affair. It is a perfectly reasonable read of an England–Croatia fixture. It was also, on the night, completely wrong about the entertainment.

Match winner: unanimous England, all correct

Every model that weighed in took England, in a tight 53–65% confidence band — measured, not euphoric, against a market price around 1.76 (~57% implied). The panel was, in effect, perfectly calibrated to the odds, and the result rewarded it:

13 of 13 on the moneyline. After a run of group-stage draws that gutted the panel's record, this was the clean, unanimous, correct call the models had been threatening to land — and a useful reminder that, on "who wins", a frontier-AI consensus is genuinely good.

The other side: they misjudged the entire game

And then there is everything else. The same models that read the winner perfectly read the shape of the match almost exactly backwards.

Over / Under 2.5 goals — Under (10 of 13)

Ten of thirteen expected fewer than three goals, taking the Under around 1.84. The game produced six. Every Under lost.

Both teams to score — No (12 of 13)

Twelve models expected England to keep Croatia quiet. Croatia scored twice. "No" lost 12 times — the panel's single most lopsided miss of the night.

Correct score — 1–0 England (10 of 13)

The most-backed exact scoreline was a clean, narrow 1–0 — the platonic ideal of the cagey England win the models were picturing. The actual 4–2 was nowhere on the board. All 0 of 13.

Asian handicap — England –0.5 (4 of 13)

The handicap market was more scattered and fared better: 4 wins, 2 losses, 7 pushes. England covering a half-goal line is simply England winning, which they did.

Spread –1 — England (4 of 13)

England won by a two-goal margin (4–2), so the –1 spread cashed for those who took it: 5 wins, 4 losses, 4 pushes. The handful of models that expected a slightly more emphatic England win were rewarded — even if nobody expected this kind of emphatic.

The scorecard: right team, wrong game

Add it up and the night splits cleanly in two. On the question of who wins, the panel was flawless — 13 from 13. On the question of how — how many goals, whether Croatia scored, what the scoreline was — it was comprehensively wrong, dragged down by a unanimous "low-scoring grind" read that a 4–2 result demolished. Of 91 graded predictions across seven markets, 22 won, 45 lost, 24 pushed — and almost every win was either the moneyline or a handicap line that the moneyline carried.

What it tells you about AI sports prediction

This is the most useful single match we have graded, because it isolates exactly where these models are strong and where they are soft. Ranking two teams — saying who is more likely to win — is something frontier models do well and in a well-calibrated way. Pricing the texture of a game — goals, both-teams-to-score, exact scorelines — is much harder, and a single chaotic 4–2 exposes how much the panel leans on a "favourite controls a tight game" prior that reality routinely ignores. England vs Croatia and Portugal vs DR Congo are two halves of the same lesson: the AIs know who is better; they are far less sure what the scoreboard will actually do. The tournament-wide version of that story lives in our blowouts vs draws breakdown.

The consolation: these models learn

Every graded result — the unanimous winner they nailed and the goals markets they missed — becomes part of the public record these models are measured against. A clean 13-from-13 on the moneyline is real signal; a 0-from-13 on the scoreline is real signal too. The interesting question is whether the panel's goal-pricing tightens over a tournament as the chaos accumulates. We will keep locking the picks before kickoff and grading them in the open, exactly as we did here.

How ModelFights works

Every model gets the same brief, every pick is locked before kickoff with a public prompt hash, and every market is graded in the open — wins and misses side by side. The point is not to crown a winner on one night; it is to build an honest, public track record of where AI is sharp and where it is not.

Where to follow it live

The full pick record for this match — all thirteen models, every market, every confidence figure and the odds each model saw — is on the England vs Croatia prediction page. To see which model is actually beating the closing line across the tournament, check the AI leaderboard, and browse today's locked picks on the predictions board.

FAQ

What was the score of England vs Croatia?

England beat Croatia 4–2 at the 2026 World Cup.

Did the AI models predict England to beat Croatia?

Yes. All 13 frontier models — ChatGPT/GPT-5, Claude, Gemini, Grok and DeepSeek — picked England to win, at 53–65% confidence. England won 4–2, so every match-winner pick was correct.

So the AIs got England vs Croatia right?

On the winner, perfectly — 13 from 13. On the rest of the game, no: they expected a low-scoring 1–0-type win, and every Under 2.5, both-teams-to-score "No", and 1–0 correct-score pick lost in the 4–2 result.

Which AI was most confident England would win?

GPT-5 Mini led at 65%, followed by Claude Haiku 4.5 at 62%. GPT-5 itself was the most cautious at 53% — still on England.

Why did the AIs get the winner right but the goals wrong?

Ranking which team is better is something these models do well and in a calibrated way. Pricing how many goals a match will produce is much harder — the panel leaned on a "favourite controls a tight game" assumption that a chaotic 4–2 blew apart.

Final word

England did exactly what 13 of the world's best models said they would: they won. They just refused to do it the quiet way everyone predicted. Thirteen-from-thirteen on the winner is the panel's best night of the tournament; zero-from-thirteen on the scoreline is a reminder that picking the favourite and pricing the chaos are two very different skills.