Round by round. Twenty forecasters.
Each model re-forecasts the FIFA World Cup 2026 bracket after every stage, using completed matches as new context. Below is what every model called at the end — champion, runner-up, third place, top scorer — and how accurate those calls turned out to be over all 104 matches.
| # | Model | Accuracy | Champion | Runner-up | Third | Top scorer |
|---|---|---|---|---|---|---|
Llama 3.3 70B | 74.9% | ARG | ESP | FRA | Lionel Messi9 | |
Claude Opus 4.8 | 74.1% | ESP | ARG | FRA | Kylian Mbappe9 | |
GPT-5.5 | 74.1% | ESP | ARG | FRA | Lionel Messi8 | |
04 | Claude Haiku 4.5 | 73.6% | ARG | ESP | FRA | Lionel Messi9 |
05 | GPT-5 Mini | 73.5% | ESP | ARG | FRA | Kylian Mbappe8 |
06 | Qwen 3.7 Max | 73.5% | ESP | ARG | FRA | Lionel Messi8 |
07 | Gemini 2.5 Flash | 73.1% | ARG | ESP | ENG | Lionel Messi12 |
08 | Llama 4 Maverick | 73.1% | ESP | ARG | FRA | Lionel Messi8 |
09 | Gemini 2.5 Pro | 73.0% | ESP | ARG | FRA | Kylian Mbappe8 |
10 | Qwen 3.7 Plus | 73.0% | ESP | ARG | FRA | Lionel Messi8 |
11 | Claude Sonnet 4.6 | 72.9% | ESP | ARG | FRA | Erling Haaland8 |
12 | Mistral Large | 72.8% | ESP | ARG | FRA | Lionel Messi9 |
13 | Grok 4.20 | 72.5% | ESP | ARG | FRA | Lionel Messi12 |
14 | DeepSeek V4 Pro | 72.2% | ESP | ARG | FRA | Kylian Mbappe8 |
15 | Perplexity Sonar | 72.1% | ARG | ESP | FRA | Lionel Messi9 |
16 | DeepSeek R1 0528 | 71.9% | ARG | ESP | FRA | Kylian Mbappe9 |
17 | Moonshot Kimi K2.6 | 71.4% | ESP | ARG | FRA | Kylian Mbappe8 |
18 | Gemini 2.5 Flash Lite | 71.0% | ARG | ESP | FRA | Kylian Mbappe9 |
19 | GPT-5 Nano | 70.8% | ESP | ARG | FRA | Kylian Mbappe9 |
20 | Mixtral 8x22B | 69.2% | ARG | ESP | FRA | Lionel Messi11 |
Accuracy over time
Mean match-prediction score for each round, as real results accumulated in the prompts. The quarterfinals are the cohort's sharpest round — reality delivered the models' own favourite scorelines — and the semifinals its weakest, where most of the field backed France over eventual champion Spain. The third-place match is excluded as an outlier. Hover a legend chip to isolate a model.
Group stage · real results
Click any group to see how the 20 models predicted each match.
Knockout bracket · real results
Click any match to compare all 20 model predictions with the actual outcome.