Experiment 05a · Final Test · With History
One final, ten tries. How much does each model disagree with itself?
25 AI models each simulated the WC26 Final ten times. Same match, ten independent runs — so we can measure the variance in each model's scorelines, scorers, assists and MVPs.
Real final
🇪🇸Spain1
🇦🇷Argentina0
Best model
Gemini 2.5 Pro
93.9%
Worst model
Llama 3.3 70B
33.1%
Avg accuracy
70.1%
25 × 10 runs
Leaderboard
25 models · sorted by accuracy| # | Model | xG 🇪🇸 | xG 🇦🇷 | Modal score | W / D / L % | Top scorer | MVP | Accuracy |
|---|---|---|---|---|---|---|---|---|
| 1 |  Gemini 2.5 Pro Google | 1.60 | 0.50 | 2–150% | 100/0/0 | Mikel Oyarzabal (0.9/g) | Rodri 100% | 93.9% |
| 2 |  GPT-5.6 Sol OpenAI | 1.50 | 0.50 | 1–050% | 100/0/0 | Mikel Oyarzabal (0.9/g) | Rodri 90% | 92.3% |
| 3 |  Claude Opus 4.8 Anthropic | 2.00 | 1.00 | 2–1100% | 100/0/0 | Mikel Oyarzabal (1.0/g) | Lamine Yamal 80% | 89.1% |
| 4 | Claude Fable 5 Anthropic | 2.00 | 1.00 | 2–1100% | 100/0/0 | Mikel Oyarzabal (1.0/g) | Lamine Yamal 80% | 89.1% |
| 5 | GPT-5.6 Luna OpenAI | 2.00 | 1.00 | 2–1100% | 100/0/0 | Mikel Oyarzabal (1.0/g) | Mikel Oyarzabal 60% | 89.1% |
| 6 | GPT-5.6 Terra OpenAI | 1.50 | 0.70 | 2–150% | 80/20/0 | Mikel Oyarzabal (1.0/g) | Rodri 40% | 87.9% |
| 7 | GPT-5.5 OpenAI | 1.90 | 1.00 | 2–190% | 90/10/0 | Lionel Messi (1.0/g) | Lamine Yamal 60% | 84.7% |
| 8 |  Qwen 3.7 Plus Alibaba | 1.60 | 1.10 | 2–160% | 70/10/20 | Lamine Yamal (1.0/g) | Lamine Yamal 70% | 83.5% |
| 9 | Claude Sonnet 4.6 Anthropic | 2.00 | 1.10 | 2–190% | 90/10/0 | Mikel Oyarzabal (1.0/g) | Mikel Oyarzabal 70% | 82.4% |
| 10 | Qwen 3.7 Max Alibaba | 1.80 | 1.10 | 2–180% | 80/10/10 | Lamine Yamal (1.0/g) | Mikel Oyarzabal 40% | 81.3% |
| 11 |  DeepSeek V4 Pro DeepSeek | 1.60 | 1.10 | 2–160% | 70/10/20 | Mikel Oyarzabal (1.0/g) | Rodri 60% | 79.1% |
| 12 | GPT-5 Mini OpenAI | 1.40 | 1.30 | 2–140% | 40/30/30 | Mikel Oyarzabal (1.0/g) | Lionel Messi 50% | 72.4% |
| 13 | Gemini 2.5 Flash Lite Google | 1.40 | 1.60 | 1–260% | 40/0/60 | Mikel Oyarzabal (0.9/g) | Lautaro Martinez 60% | 69.0% |
| 14 | GPT-5 Nano OpenAI | 1.30 | 1.00 | 2–130% | 30/70/0 | Lionel Messi (0.9/g) | Lamine Yamal 50% | 68.2% |
| 15 | DeepSeek R1 0528 DeepSeek | 1.30 | 1.50 | 1–250% | 30/20/50 | Lionel Messi (1.0/g) | Lionel Messi 50% | 66.8% |
| 16 |  Grok 4.20 xAI | 1.40 | 2.00 | 1–260% | 0/40/60 | Lionel Messi (1.0/g) | Lionel Messi 100% | 66.8% |
| 17 |  GLM 5.2 z.ai | 1.50 | 1.40 | 2–140% | 40/30/30 | Lionel Messi (0.9/g) | Lionel Messi 50% | 64.5% |
| 18 | Gemini 2.5 Flash Google | 1.00 | 2.00 | 1–2100% | 0/0/100 | Lionel Messi (0.9/g) | Lionel Messi 80% | 63.4% |
| 19 |  Llama 4 Maverick Meta | 1.00 | 2.00 | 1–2100% | 0/0/100 | Lautaro Martinez (1.0/g) | Lionel Messi 100% | 63.4% |
| 20 |  Moonshot Kimi K2.6 Moonshot | 1.20 | 2.10 | 1–270% | 0/10/90 | Lionel Messi (1.0/g) | Lionel Messi 90% | 56.7% |
| 21 |  Mixtral 8x22B Mistral | 1.40 | 1.70 | 1–250% | 40/0/60 | Lionel Messi (0.9/g) | Lionel Messi 60% | 55.5% |
| 22 | Mistral Large Mistral | 1.00 | 2.00 | 1–2100% | 0/0/100 | Lamine Yamal (1.0/g) | Lionel Messi 100% | 50.0% |
| 23 |  Perplexity Sonar Perplexity | 1.00 | 2.00 | 1–2100% | 0/0/100 | Lamine Yamal (1.0/g) | Lautaro Martinez 70% | 36.5% |
| 24 | Claude Haiku 4.5 Anthropic | 1.00 | 2.00 | 1–2100% | 0/0/100 | Lautaro Martinez (1.0/g) | Lionel Messi 90% | 33.1% |
| 25 | Llama 3.3 70B Meta | 1.00 | 2.00 | 1–2100% | 0/0/100 | Lionel Messi (1.0/g) | Lionel Messi 100% | 33.1% |