Experiment 05b · Final Test · No History
One final, ten tries. How much does each model disagree with itself?
25 AI models each simulated the WC26 Final ten times, this time with the tournament match history withheld from the prompt. 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
Perplexity Sonar
95.6%
Worst model
GPT-5 Mini
55.9%
Avg accuracy
83.5%
25 × 10 runs
Leaderboard
25 models · sorted by accuracy| # | Model | xG 🇪🇸 | xG 🇦🇷 | Modal score | W / D / L % | Top scorer | MVP | Accuracy |
|---|---|---|---|---|---|---|---|---|
| 1 |  Perplexity Sonar Perplexity | 1.60 | 0.10 | 2–050% | 100/0/0 | Mikel Oyarzabal (1.0/g) | Rodri 100% | 95.6% |
| 2 |  Gemini 2.5 Pro Google | 1.10 | 0.20 | 1–080% | 90/10/0 | Mikel Oyarzabal (1.0/g) | Rodri 90% | 93.4% |
| 3 |  Moonshot Kimi K2.6 Moonshot | 2.00 | 0.20 | 2–080% | 100/0/0 | Mikel Oyarzabal (1.0/g) | Rodri 100% | 92.5% |
| 4 |  Qwen 3.7 Max Alibaba | 1.90 | 0.40 | 2–050% | 100/0/0 | Mikel Oyarzabal (1.1/g) | Rodri 70% | 92.3% |
| 5 | Qwen 3.7 Plus Alibaba | 2.00 | 0.30 | 2–070% | 100/0/0 | Mikel Oyarzabal (1.0/g) | Lamine Yamal 60% | 92.1% |
| 6 |  Llama 4 Maverick Meta | 2.00 | 0.40 | 2–060% | 100/0/0 | Lamine Yamal (1.0/g) | Lamine Yamal 100% | 91.7% |
| 7 | Gemini 2.5 Flash Lite Google | 1.60 | 0.50 | 2–150% | 100/0/0 | Mikel Oyarzabal (1.0/g) | Rodri 100% | 89.4% |
| 8 |  Claude Opus 4.8 Anthropic | 2.00 | 1.00 | 2–1100% | 100/0/0 | Mikel Oyarzabal (1.0/g) | Lamine Yamal 90% | 89.1% |
| 9 |  Mixtral 8x22B Mistral | 2.00 | 1.00 | 2–1100% | 100/0/0 | Lamine Yamal (1.0/g) | Rodri 100% | 89.1% |
| 10 | Claude Fable 5 Anthropic | 2.00 | 1.00 | 2–1100% | 100/0/0 | Mikel Oyarzabal (1.0/g) | Mikel Oyarzabal 70% | 86.9% |
| 11 |  GPT-5.6 Sol OpenAI | 1.70 | 0.80 | 2–170% | 90/10/0 | Lionel Messi (0.8/g) | Rodri 80% | 86.8% |
| 12 |  GLM 5.2 z.ai | 1.90 | 0.60 | 2–150% | 90/10/0 | Mikel Oyarzabal (1.0/g) | Rodri 70% | 86.4% |
| 13 |  DeepSeek V4 Pro DeepSeek | 1.90 | 0.60 | 2–160% | 100/0/0 | Mikel Oyarzabal (0.9/g) | Rodri 100% | 84.8% |
| 14 | GPT-5.5 OpenAI | 1.80 | 1.00 | 2–180% | 80/20/0 | Lionel Messi (0.9/g) | Rodri 70% | 84.7% |
| 15 | Gemini 2.5 Flash Google | 1.80 | 0.50 | 2–050% | 90/0/10 | Mikel Oyarzabal (1.0/g) | Rodri 90% | 84.5% |
| 16 | GPT-5.6 Luna OpenAI | 1.80 | 1.00 | 2–180% | 80/20/0 | Lionel Messi (1.0/g) | Lamine Yamal 90% | 82.4% |
| 17 | Claude Sonnet 4.6 Anthropic | 1.50 | 0.90 | 2–150% | 60/40/0 | Mikel Oyarzabal (1.0/g) | Rodri 60% | 81.3% |
| 18 | Claude Haiku 4.5 Anthropic | 1.80 | 0.60 | 2–060% | 80/0/20 | Mikel Oyarzabal (1.0/g) | Rodri 80% | 80.5% |
| 19 | Mistral Large Mistral | 1.70 | 0.90 | 2–170% | 90/0/10 | Mikel Oyarzabal (1.0/g) | Rodri 90% | 80.1% |
| 20 | GPT-5.6 Terra OpenAI | 1.40 | 1.00 | 1–160% | 40/60/0 | Mikel Oyarzabal (0.8/g) | Rodri 50% | 75.7% |
| 21 |  Grok 4.20 xAI | 1.50 | 1.40 | 2–150% | 50/10/40 | Lamine Yamal (1.0/g) | Lionel Messi 50% | 74.6% |
| 22 | GPT-5 Nano OpenAI | 1.30 | 0.90 | 0–040% | 40/60/0 | Lamine Yamal (0.9/g) | Lamine Yamal 80% | 74.5% |
| 23 | DeepSeek R1 0528 DeepSeek | 1.10 | 0.50 | 1–030% | 60/40/0 | Mikel Oyarzabal (1.0/g) | Rodri 90% | 73.4% |
| 24 | Llama 3.3 70B Meta | 1.50 | 0.80 | 2–050% | 60/0/40 | Mikel Oyarzabal (1.0/g) | Lamine Yamal 50% | 69.9% |
| 25 | GPT-5 Mini OpenAI | 1.60 | 1.70 | 2–260% | 0/90/10 | Lionel Messi (0.9/g) | Lionel Messi 70% | 55.9% |