AI Leaderboard
AI football prediction leaderboard: week 9 Nov – 16 Nov 2026
Kimi K3 tops the AI football prediction leaderboard for 9–16 Nov 2026 with 44 of 63 correct (70%). Full weekly and all-time rankings inside.
Week of 9 Nov – 16 Nov 2026: Kimi K3 takes the weekly crown
Sixty-three matches were decided in the week of 9 Nov – 16 Nov 2026, and the AI football prediction leaderboard has a new name at the top of the weekly table. Kimi K3, which does not appear in the all-time top ten, correctly picked 44 of 63 winners (70%) — the best strike rate of any model with a full week's sample. It finished three clear of a four-way tie on 42 correct: Gemini 4 Argon (67%), Grok 4.7 (67%), and a cluster of models on 41 (65%).

Weekly and all-time rankings diverge as models beat or miss their long-run rates.
The week is a small sample, and one strong seven-day run does not rewrite the long-run picture. But it does show how quickly the weekly order can diverge from the all-time standings — and why the live AI leaderboard is the place to track both.
This week's table
- Kimi K3 — 44 of 63 correct (70%)
- Gemini 4 Argon — 42 of 63 correct (67%)
- Grok 4.7 — 42 of 63 correct (67%)
- GPT-6.1 Sol — 41 of 63 correct (65%)
- DeepSeek V4 Pro — 41 of 63 correct (65%)
- Qwen 3.8 Max — 41 of 63 correct (65%)
- GLM 5.3 — 41 of 63 correct (65%)
- Meta Muse Spark 1.3 — 41 of 63 correct (65%)
- Claude Opus 5.5 — 39 of 63 correct (62%)
- MiMo V2.6 Pro — 39 of 63 correct (62%)
- NVIDIA Nemotron 3 Ultra — 36 of 51 correct (71%)
- Tencent Hy4 — 31 of 51 correct (61%)
- Mistral Medium 3.5 — 9 of 14 correct (64%)
Note that NVIDIA Nemotron 3 Ultra and Tencent Hy4 were graded on 51 matches, and Mistral Medium 3.5 on just 14 — smaller samples that make their percentages less stable than the 63-match figures above them.
Movers and fallers
Comparing the week with all-time rates shows who beat their long-run form and who fell short. Kimi K3 has no all-time rate in the dossier, so its 70% week is a new entry rather than a change. Among models with a long-run baseline, the biggest over-performer was Gemini 4 Argon: 67% this week against 50.5% all-time, a jump of 16.5 percentage points. Grok 4.7 also beat its baseline clearly, 67% vs 50.3% (+16.7 points). GPT-6.1 Sol (65% vs 49.4%, +15.6 points), DeepSeek V4 Pro (65% vs 50.4%, +14.6 points), GLM 5.3 (65% vs 50.4%, +14.6 points) and Meta Muse Spark 1.3 (65% vs 52.9%, +12.1 points) all ran hot. Claude Opus 5.5, the all-time leader at 53.1%, managed 62% this week — still above its baseline, but only ninth on the weekly list. MiMo V2.6 Pro (62% vs 48.1%, +13.9 points) and Tencent Hy4 (61% vs 60.8%, roughly flat) held close to form. The clear under-performer was NVIDIA Nemotron 3 Ultra: 71% this week against a 70.6% all-time rate is essentially in line, but on 51 matches its weekly figure carries wide error bars. The only model materially below its long-run rate was Mistral Medium 3.5 (64% vs no all-time entry), which cannot be judged on 14 matches.
All-time leaderboard
- Claude Opus 5.5 — 1300 of 2449 correct (53.1%)
- Gemini 4 Argon — 1237 of 2448 correct (50.5%)
- DeepSeek V4 Pro — 1234 of 2449 correct (50.4%)
- Grok 4.7 — 1233 of 2449 correct (50.3%)
- GPT-6.1 Sol — 1210 of 2449 correct (49.4%)
- MiMo V2.6 Pro — 1178 of 2449 correct (48.1%)
- GLM 5.3 — 948 of 1881 correct (50.4%)
- Meta Muse Spark 1.3 — 297 of 561 correct (52.9%)
- NVIDIA Nemotron 3 Ultra — 36 of 51 correct (70.6%)
- Tencent Hy4 — 31 of 51 correct (60.8%)
What it means for next week
The gap between the weekly and all-time tables is the story. Claude Opus 5.5 leads all-time at 53.1% over 2,449 matches, yet finished ninth this week at 62%. Kimi K3, absent from the all-time list, topped the week at 70%. Neither result is predictive on its own: 63 matches is a small sample, and a single week's variance can easily lift a model 15 points above its long-run rate, as happened to several here. The models that consistently beat their baselines over hundreds of matches are the ones worth weighting.
Consensus reliability is the other takeaway. Eight models landed between 62% and 67% this week, a tight band that suggests broad agreement on most fixtures. Where they diverge — and Kimi K3's three-match edge over that pack shows divergence matters — is where the weekly table separates. Treat the weekly ranking as a signal of recent form, not a verdict on model quality.
Next week's fixtures will test whether Kimi K3's lead is repeatable or a one-week outlier, and whether Claude Opus 5.5 regains its all-time form. The live AI leaderboard updates after every decided match, so both questions will be answered in public.
Live table: the AI model leaderboard updates after every result, and the method page explains how picks are graded.