World Cup 2026 AI Predictions: What the Models Got Right (and Wrong)

Key takeaways

Try our awesome AI for free

Ask me anything...

Spain won the 2026 World Cup (1-0 over Argentina after extra time, on July 19, 2026). Before the tournament, the press asked the AIs: Claude and Perplexity said Spain, ChatGPT answered Spain or France depending on the survey, Mistral picked France or Brazil, DeepSeek and Qwen bet on Argentina. The same model gave different answers depending on the outlet and the prompt: a conversational AI does not compute a probability, it generates plausible text. For its part, the Opta supercomputer had Spain as favorite at just 16.1%: the correct final four, but 83.9% of its 25,000 simulations crowned a different champion. Here is the full verdict, prediction by prediction.

• Final: Spain 1, Argentina 0 after extra time, goal by Ferran Torres, on July 19, 2026 at MetLife Stadium.
• Conversational AIs answered Spain, France, Brazil or Argentina depending on the model… and on the survey: the same ChatGPT picked Spain for the AFP and France for 01net.
• Biggest collective miss: Brazil in the semi-finals for all five AIs surveyed by 01net, knocked out in the Round of 16 by Erling Haaland’s Norway.
• Nicest feat: Opta and probwin’s Elo model both had Spain as favorite, with an exact final four for Opta.
• The AI that is actually useful for soccer structures your match analysis; it does not predict scores, and no forecast guarantees a win.

The 2026 World Cup ended on July 19, 2026, yet searches for “World Cup 2026 AI predictions” are still alive: simulators, retrospectives, prediction reviews. No wonder: the 2026 edition was a real-scale experiment. For the first time, a World Cup with 48 teams and 104 matches, co-hosted by the United States, Canada and Mexico from June 11 to July 19, 2026, played out while conversational AIs had become everyday objects. Newsrooms noticed: before kickoff, 01net, Le Parisien, the newswire AFP picked up by Libération and Konbini (French-language outlets) fed the full bracket to the models, while sports supercomputers fine-tuned their millions of simulations.

The result: a corpus of dated, signed and verifiable predictions that the real tournament put through the wringer. That is exactly what this article does: it confronts every prediction published before or during the 2026 World Cup with what actually happened, then draws the useful lesson about what an AI can (and cannot) do when you ask it for a forecast. Beyond soccer, we had already run this exercise for horse racing in our article on AI horse racing predictions; the World Cup offers a far bigger stage, and sharper lessons.

Before kickoff: what each AI answered

The opening match, on June 11, 2026, saw Mexico beat South Africa 2-0. In the days before, press surveys had asked the AIs to predict the entire competition. The most complete one is 01net’s, published on June 11, 2026: five models surveyed, each with the full bracket, the finalists and the predicted top scorer.

Claude’s verdict fit in one sentence reported by 01net: “It is Spain, described as the best team in the world right now, that would win the competition” (translated from the French). ChatGPT justified its French confidence through Kylian Mbappé’s track record, “already winner of the 2022 World Cup Golden Boot with 8 goals”, noting that “bookmakers currently place him as the main favorite for the 2026 Golden Boot”, and that “France is one of the teams most likely to reach the final” (translated from the French). Perplexity even had France eliminated by Argentina in the Round of 16, before a Spain vs Brazil final.

One detail is delightful in hindsight: to reach its France vs Spain final, ChatGPT went through semi-finals of Brazil vs Spain and France vs Norway. Norway, precisely the team that would knock Brazil out in the Round of 16.

The geographic split of the predictions

The AFP survey, published on June 12 and 13, 2026 and picked up by Libération under the headline “France for Mistral, Spain for Claude and ChatGPT… AI gets into the forecasting game” (in French), then by Konbini, reveals a striking split. Konbini’s summary: “Claude and OpenAI predict a Spain victory, Mistral leans toward Les Bleus, DeepSeek and Qwen bet on Argentina” (translated from the French). The two Chinese models surveyed by the specialist site Decrypt both bet on Lionel Messi’s Argentina, the defending champion. Meanwhile, Bank of America analysts told the AFP they gave the same mission to Copilot, Microsoft’s assistant, which “places Spain on a par with France” (translated from the French); and Tom’s Guide reached the same trio after quizzing Gemini, ChatGPT and Perplexity: Spain on top, France just behind.

The same model, two different answers

That is the most instructive point of the whole collection. According to the AFP, picked up by Libération and Le Parisien, ChatGPT made Spain champion. According to 01net, which asked the question two days earlier with its own simulation grid, ChatGPT made France champion. Same for Mistral: France for the AFP, Brazil for 01net. Gemini: Spain for Tom’s Guide, France for 01net.

These divergences are not reporting errors: they show that a language model does not “know” the future of the tournament. It generates a plausible answer from its training corpus, and that answer changes with the wording of the question, the requested format (a full sheet with semi-finals and scorers forces the model to commit to details) and the model’s internal random draw. Put differently: asking a conversational AI for a forecast is rolling a very well-written die.

What actually happened

The real script delivered a dream final four and a final of rare tactical discipline. In chronological order, the moments that swung the tournament:

• June 11: Mexico beat South Africa 2-0 to open the competition.
• July 5: Brazil, whom all five of 01net’s AIs sent to the semi-finals, was eliminated in the Round of 16 by Norway 1-2, on an Erling Haaland brace, despite a Neymar penalty in the tenth minute of stoppage time. Neymar then announced his international retirement (“It’s over”, Reuters, July 6, 2026).
• July 9 and 10: in the quarter-finals, England edged Mexico 3-2; Opta’s four pre-tournament favorites (Spain, France, Argentina, England) were all still alive, while the teams ranked fifth to eighth (Portugal, Brazil, Germany, Netherlands) had all vanished.
• July 14 and 15: semi-finals. France fell 0-2 to Spain; Argentina came back against England 2-1 on a Lautaro Martínez goal in the 92nd minute, after two Lionel Messi assists.
• July 18: third-place playoff in Miami. England led 4-0 at half-time and eventually won 6-4 against France, in a ten-goal game. During that match Kylian Mbappé became the all-time top scorer of the World Cup.
• July 19: final at MetLife Stadium (New Jersey). Spain beat Argentina 1-0 after extra time, with a Ferran Torres goal early in the second period of extra time, against an Argentina down to ten after Enzo Fernández was sent off for two yellow cards. A second star for La Roja, sixteen years after 2010.

Two notes for prediction duel enthusiasts. Harry Kane finished with 6 goals and became England’s all-time top World Cup scorer (14 in total, ahead of Gary Lineker’s 10). And Lamine Yamal, whom Claude sent to the Golden Boot with 8 goals, scored only one, against Saudi Arabia; he did attempt 49 dribbles, more than any other player in the tournament according to the BBC.

The prediction tribunal: filter the 13 forecasts

We gathered in the filter below the predictions published before the tournament (press surveys) and the statistical models’ forecasts, each one dated and sourced. Check the conditions to keep only the predictions you care about: conversational AIs, statistical models, those that published a number, those that named the right champion. The counter updates on every click.

What the filter reveals

Checking only “Winner found” leaves six cards: two press answers from ChatGPT and Gemini, Claude, Perplexity, and the two statistical models. Checking “Conversational AI” and “Published number” shows that almost every numbered grid came from the LLMs themselves, including the scorer bets. Three lessons stand out:

• Brazil was the collective trap. All five of 01net’s AIs sent Brazil to the semi-finals, Mistral crowned it, and it went out in the Round of 16 against Haaland’s Norway. “Eternal favorite” Brazil weighs heavily in training corpora: that is a memory bias, not analysis.
• The numbers beat the verdicts. Mbappé announced at 9 goals (ChatGPT and Gemini) against 10 real ones, Harry Kane at exactly 6 goals (ChatGPT and Gemini): on statistical volumes, the models were remarkably sharp, sharper than on the champion’s name.
• Spain divided the AIs… and half-united them. Claude and Perplexity made it champion, ChatGPT and Gemini saw it losing the final to France, DeepSeek and Qwen preferred its final opponent. No consensus, therefore no science.

Why the same AI answers differently

A language model like ChatGPT, Claude, Gemini, Mistral, DeepSeek or Qwen contains no sports simulation engine. It has read millions of texts (articles, standings, fan debates) and, when you ask it for a winner, it produces the statistically most plausible continuation of the current text. That is why:

• The answer depends on the question. “Who will win?” invites a single name (often the most-cited favorite in the corpus); “simulate the whole bracket” forces the model to invent semi-finals, multiplying the divergence points.
• The answer depends on the draw. Models generate with a dose of randomness (the “temperature”): two consecutive sessions can pick Spain then France, with neither being a “calculation”.
• The corpus outweighs actual form. Spain, the 2024 European champion, dominated pre-tournament coverage; Brazil carries a five-star aura; Mbappé’s France feeds the French-language press… and the French-language model adds its own bias.
• Data stops at the training cutoff. Injuries, groups, June 2026 form: a model without web search reasons over an idealized version of soccer, not the real edition.

This is not a hidden flaw, it is the nature of the object. Researchers at LMU Munich even launched LLM SoccerArena, a public site that benchmarks models’ predictive accuracy on real matches. One of them, Stefan Feuerriegel, summed it up for the AFP: “We need benchmarks that do not only evaluate abstract tasks, but how models handle evolving information, uncertainty, and outcomes that can then be compared” with real results (translated from the French).

In other words: the question is not “which AI predicts best” but “how to honestly measure what these models can guess”. The 2026 World Cup gave a clear answer: on a champion’s name, they diverge; on volumes (a named striker’s goal count), they get close; on Norway’s run, a real quarter-finalist that only ChatGPT saw… in the semi-finals, nobody nailed it.

Did the statistical models do better?

Yes, and above all: their hit was better earned. The Opta supercomputer, which powers The Analyst, had run 25,000 tournament simulations before kickoff: Spain won 16.1% of them, ahead of France, Argentina and England. Come the quarter-finals, those four teams were exactly the four still standing, while Portugal, Brazil, Germany and the Netherlands, ranked fifth to eighth, had all been eliminated. The pre-tournament hierarchy survived to the end: remarkably rare at that level of precision.

But Opta’s real lesson lies elsewhere: 16.1% means that in 83.9% of its 25,000 simulated worlds, Spain did not win. The model was “right” while simultaneously proclaiming that the most likely outcome was a Spanish defeat. That is the whole difference between a forecast and a probability: the statistical model quantifies its uncertainty, an LLM’s answer masks it. For its part, the Elo model featured by probwin.com gave Spain a 20.7% probability of the title, ahead of Argentina (15.7%), relying on the best Elo rating in the field (2157): same hierarchy, same owned uncertainty, same correct final verdict.

The box that should come before any bet

What an AI is actually good for in soccer

If the raw forecast is not a science, AI remains an excellent second brain for analysis. What it can honestly do: gather and cross information (recent form, head-to-head history, squads, absences), lay it out in a table, summarize a tournament, prep a debate between friends or a fan-club piece. What it cannot do: turn that analysis into an exact score or a certainty.

It all lives in the prompt. A raw prediction question gets an optimistic fan’s answer; a structured analysis brief gets a usable dossier.

The second prompt is exactly the kind Nation AI, the AI assistant of nation.ai handles best: text, structure, comparison, with a free trial and no sign-up. Nation AI does not predict scores, it structures your match analysis: which is already what the best models do, better targeted. And if you want to systematize this kind of brief for sports or any other topic, our prompt generator builds the skeleton for you; all our tools are gathered on the AI tools page.

FAQ about the 2026 World Cup AI predictions

Conclusion: the tribunal’s verdict

The 2026 World Cup ruled: Spain champions, Brazil gone too soon, Mbappé the all-time scorer, and a banquet of half-digested predictions. The conversational AIs had fine instincts (Spain for several of them, Mbappé’s 9 then 10 goals, Kane’s 6 goals to the digit) and one monumental collective miss (Brazil in the semi-finals). Above all, they showed their true nature: the same model answers different things to different outlets. An AI forecast is not a measurement, it is a very well-written opinion.

The right way to use AI around soccer is to treat it as an analyst, not an oracle. And for that, you need to talk to it properly.