Open benchmark · every prediction committed before the event
Which model predicts reality best?
Claude Fable 5.1 leads the Football leagues board with 55.9% Right outcome — but Kimi K3 is only 1.4% behind, which is inside the measurement error: first place goes to Claude Fable 5.1 in 53% of resampled runs and to Kimi K3 in 35%.
Right outcome · Football leagues
no clear leaderClaude Fable 5.1 called 55.9% of matches right
Places 1 and 2 are close enough that the order may still flip.
351 of 424 events resolved · 23 models · ranked by Right outcome · Every prediction was published before the event. The models had no internet access.
| Model | Right outcome | Exact | Points/event | Rank (90%) | P(#1) |
|---|---|---|---|---|---|
| 1 Claude Fable 5.1 Anthropic | 55.9% | 9.5% | 1.44 | 1–4 | 53% |
| 2 Kimi K3 Moonshot AI | 54.5% | 9.9% | 1.41 | 1–4 | 35% |
| 3 Gemini 3.5 Flash Google | 51.6% | 9.9% | 1.33 | 1–4 | 18% |
| 3 GLM-5.2 Z.ai | 51.6% | 12.7% | 1.38 | 1–5 | 9% |
| 5 DeepSeek V4 Pro DeepSeek | 50.0% | 10.2% | 1.30 | 4–6 | 0% |
| 5 Qwen 3.7 Max Alibaba | 50.0% | 11.1% | 1.32 | 2–6 | 2% |
| – Gemini 4 Hydrogen Google 1/351 events — not ranked | 100.0% | 0.0% | 2.00 | — | — |
| – GPT-6.1 Sol OpenAI 150/351 events — not ranked | 60.7% | 12.7% | 1.57 | — | — |
| – GLM 5.3 Prime Z.ai 61/351 events — not ranked | 60.7% | 14.8% | 1.62 | — | — |
| – Claude Opus 5.5 Anthropic 150/351 events — not ranked | 59.3% | 10.0% | 1.53 | — | — |
| – Grok 4.7 xAI 150/351 events — not ranked | 58.7% | 11.3% | 1.51 | — | — |
| – Qwen3.8 2.4t A95b Alibaba 103/351 events — not ranked | 58.3% | 10.7% | 1.50 | — | — |
| – Qwen3.8 27b Alibaba 102/351 events — not ranked | 56.9% | 12.7% | 1.49 | — | — |
| – GLM 5.3 Flashx Z.ai 103/351 events — not ranked | 56.3% | 12.6% | 1.46 | — | — |
| – Gemma 4 31b It Google 113/351 events — not ranked | 54.9% | 14.2% | 1.47 | — | — |
| – Kimi K2.7 Code Moonshot AI 103/351 events — not ranked | 54.4% | 11.7% | 1.41 | — | — |
| – GPT-5.5 OpenAI 137/351 events — not ranked | 51.8% | 10.2% | 1.36 | — | — |
| – Claude Fable 5 Anthropic 137/351 events — not ranked | 48.2% | 8.8% | 1.23 | — | — |
| – Gemma 4 26b A4b It Google 113/351 events — not ranked | 46.0% | 7.1% | 1.12 | — | — |
| – GPT-5.6 Sol OpenAI 38/351 events — not ranked | 44.7% | 5.3% | 1.11 | — | — |
| – Grok 4.6 xAI 29/351 events — not ranked | 41.4% | 6.9% | 1.07 | — | — |
| – Qwen3.8-Max Alibaba 29/351 events — not ranked | 37.9% | 10.3% | 1.07 | — | — |
| – Claude Fable 5.2 Anthropic 1/351 events — not ranked | 0.0% | 0.0% | 0.00 | — | — |
| ø Always home win reference | 43.9% |
Scroll sideways for more columns
Separable at 95%: 2 of 15 model pairs (Claude Fable 5.1–DeepSeek V4 Pro, Kimi K3–DeepSeek V4 Pro). Every other ordering in this table is within noise.
Rank intervals and P(#1) come from 4,000 paired resamples of the 179 events, seed 20260727 — same draw for every model, so the comparison stays fair. Events share context, which makes these intervals lower bounds.
The most common result in professional football, written down on 29 July 2026 — before any league event had resolved. A genuine a-priori reference: it was fixed first, the outcomes came later.
78 events open · next lock Nov 16, 2026, 00:00 UTC · 80 of 390 predictions committed
Cross-category skill — locked
A cross-category number needs at least 2 qualified categories. 1 qualifies today, so none is computed — an average over a single category would just be that category’s table under a grander name.
Not qualified yet
A category qualifies with: real data, a clean integrity check, at least 20 resolved events, at least 3 models scored on at least 80% of the shared events, and a reference written down beforehand.
- Football leagues
- too few models scored on the shared events
- Crypto
- too few models scored on the shared events
- Stock index
- too few models scored on the shared events
- Elections & politics
- sample data
- Mixed sports
- sample data
Data as of Nov 18, 2026, 14:31 UTC · 4 live categories, 2 sample · 753 events tracked · page built Nov 18, 2026, 14:32 UTC
Proof · Models · Methodology