Which AI predicts the World Cup best? We tested ChatGPT, Claude, Gemini, Copilot and Perplexity
Five AI models predicted every World Cup match. Three choose Spain, and in 28 games they are the strongly disagree.
ChatGPT, Claude, Gemini, Copilot and Perplexity recorded all 104 matches in advance. We will keep live track of who comes out on top.
- 3 out of 5 Models tip Spain as world champion. Gemini picks Brazil, Copilot picks Argentina.
- Morocco is the most popular surprise: two models will see the semi-finals, consistent with the 2022 World Cup.
- All five brackets are internally correct: no model tips a top scorer from a country that it itself eliminates early.
- 7 matches with a solo deviation: moments when exactly one model goes against the four others. Six out of seven are equals.
- All predictions are recorded before 11 June and are no longer editable. A live scoreboard counts points per match played.
Bureau Onder captured full tournament predictions from five major AI language models before the start of the 2026 FIFA World Cup. Each model was given the exact same structured assignment: predict all 72 group matches with exact results, the full knockout brackets, a top scorer top five, a dark horse and a flop. The assignment included a consistency requirement: a model's dark horse must match its own bracket. This makes the outcomes for each model comparable and verifiable.
Spain, but no unanimous verdict
ChatGPT, Claude and Perplexity tipping Spain as world champion. Gemini chooses Brazil and Copilot Argentina. France is in the semi-finals in all five brackets, but will not become champion by any model.
| Model | World champion | Losing finalist | Top scorer | Dark horse | Biggest flop |
|---|---|---|---|---|---|
| ChatGPT | Spain | Argentina | Haaland (8×) | Norway | Brazil |
| Claude | Spain | Argentina | Mbappé (8×) | Morocco | Brazil |
| Gemini | Brazil | Spain | Mbappé (7×) | US | Germany |
| Copilot | Argentina | Spain | Messi (6×) | Morocco | Netherlands |
| Perplexity | Spain | Argentina | Mbappé (7×) | Colombia | Germany |
Source: Bureau Onder - Forecasts by model, set 8 June 2026
Two models tip Brazil as the biggest disappointment, two tip Germany and one the Netherlands. Gemini is the only model that sees Brazil as champion and the only one that tips Germany as flop.
Seven times one model faces four others
In seven matches, exactly one model deviates from the four others. These are the duels that will determine the rankings: if the deviator is proven right, it grabs a lead that the rest cannot compensate for. Six of the seven solo deviations are ties.
| Model | Match | Deviant forecast | What the other four chose |
|---|---|---|---|
| Gemini | Czech Republic vs Mexico (A) | 1-1 draw | Mexico wins (4×) |
| Perplexity | Canada vs Qatar (B) | 1-1 draw | Canada wins (4×) |
| Claude | Brazil vs Morocco (C) | 1-1 draw | Brazil wins (4×) |
| Perplexity | Australia vs Turkey (D) | 1-1 draw | Turkey wins (4×) |
| Claude | Paraguay vs Australia (D) | 1-1 draw | Paraguay wins (4×) |
| Copilot | Ecuador vs Germany (E) | 1-1 draw | Germany wins (4×) |
| Perplexity | Ivory Coast vs Ecuador (E) | 2-0 Ivory Coast | Ecuador win or draw |
Source: Bureau Onder - Comparison of forecasts by model, 8 June 2026
Perplexity has the most solo deviations (3), followed by Claude (2). ChatGPT has none: the model does not deviate from the group as the only one anywhere.
Five models, five recognisable styles
The style differences below are derived from the 360 captured predictions themselves. They are verifiable in the raw data file.
ChatGPT follows the favourite. Zero solo aberrations, no prediction against a clear favourite.
Claude picks the lowest results. Predicts the most draws of the five, especially in matches between equivalent teams.
Gemini goes for goals. Predicts the highest results of the bunch, such as Brazil 5-0 Haiti and New Zealand 0-4 Belgium.
Copilot more often gives the underdog a point. Chooses only Messi as top scorer, in line with his title choice for Argentina.
Perplexity dares to stand alone. Most often the only deviation (3 solo calls) and is the only model to see an outsider like Ivory Coast win over Ecuador.
Are the brackets correct internally?
A prediction can only be taken seriously if it is internally correct: the dark horse must actually get far in its own bracket, and the top scorer must play for a country that has been in the tournament long enough to reach that number of goals. We checked this for all five models.
| Model | Dark horse | In its own bracket to | Top scorer | Country in its own bracket | Is internally correct? |
|---|---|---|---|---|---|
| ChatGPT | Norway | Semifinals | Haaland | Norway, semi-final | Yes |
| Claude | Morocco | Semifinals | Mbappé | France, semi-final | Yes |
| Gemini | US | Quarterfinals | Mbappé | France, semi-final | Yes |
| Copilot | Morocco | Semifinals | Messi | Argentina, champion | Yes |
| Perplexity | Colombia | Quarterfinals | Mbappé | France, semi-final | Yes |
Analysis: Bureau Onder, 8 June 2026
All five models deliver an internally throbbing whole. The tightest line of reasoning is that of Copilot: Argentina as champion, Messi as top scorer and a final where both meet. The least daring is that of Gemini: the US as a surprise to the quarterfinals is a safe bet for a host country with home advantage.
Morocco favourite among models
Each model picked one team that advanced beyond most people's expectations. Morocco is named twice, both times to the semi-finals.
- ChatGPT - Norway (until semi-final). Haaland as defining factor.
- Claude - Morocco (until the semi-finals). Proven in 2022, strong collective.
- Gemini - US (until the quarterfinals). Home advantage as host nation.
- Copilot - Morocco (until the semi-finals). Experience plus low expectations.
- Perplexity - Colombia (until the quarter-finals). Strong in form, underrated.
Source: Bureau Onder - Dark horse forecasts by model, 8 June 2026
Current standings: which AI model will win the pool?
Per match the prediction of each model alongside the actual result, the current ranking and the percentage of correct predictions per model. Scores are updated after each match played.
The Scorito points system
We use the Scorito points system, known from the World Cup pools in which millions of Dutch people participate. Each round counts more: a correct final winner yields six times as much as a proper group match.
| Performance | Points |
|---|---|
| Exact result correct in group stage | 45 points |
| Correct outcome but not exact in group stage | 30 points |
| Exact result correct in the Round of 16 | 45 × 1,5 |
| Exact result correct in quarterfinals | 45 × 2 |
| Exact result correct in semi-final | 45 × 3 |
| Exact result correct in final | 45 × 6 |
Source: Scorito - Standard World Cup points system
A structured master assignment, not an isolated question
All five models were given exactly the same assignment with the official FIFA playing schedule and the requirement to make their methods explicit: what factors they factor in (recent form, quality of selection, FIFA ranking, injuries, tournament context), how heavily each factor counts and what sources they use. The assignment included a consistency requirement: dark horse and flop must match their own bracket. We recorded the first generated output per model, without iterations or adjustments.
What we do not claim
- Whether and how each model consulted actual web data during prediction varies by model and session. We make no claims about this that we did not capture on a session-by-session basis.
- Injuries, suspensions or team changes after 8 June 2026 are not included.
- Models may arrive at different results with the same command when repeated. The recorded forecast is one run per model.
- The scoring system rewards exact results heavily. Models that structurally predict conservatively may suffer.
- This study measures prediction accuracy in one tournament and says nothing about the overall quality of a model for other tasks.
- World Cup playing schedule: FIFA.com, official 2026 World Cup playing schedule
- Grouping 2026 World Cup: FIFA.com, final draw December 2025
- Scoring system: Scorito, standard World Cup points system
- Predictions by model: Bureau Onder, fixed on 8 June 2026
On 19 July 2026, we will know
Three models tip Spain, two see another champion, and in seven matches one model faces the rest. On 19 July 2026, we will know which model read the World Cup best.
Dutch
English