AI football prediction leaderboard: week 5 Oct – 12 Oct 2026

Week 5 Oct – 12 Oct 2026 produced 103 decided matches across the competitions tracked by TuringStats. Ten frontier models were graded on winner picks, and GLM 5.3 was the model of the week: 59 of 103 correct (57%). That is the highest tally in the weekly table and comfortably above the model's all-time rate. Two other systems, Cohere Command R+ (08-2024) and Llama 3.1 70B Instruct, also appear with 43 decided matches each, but they are not part of the main ten-model set and their weekly percentages are noted separately below.

This week's table

1. GLM 5.3 — 59 of 103 correct (57%)
2. Claude Sonnet 5 — 55 of 103 correct (53%)
3. Gemini 3.8 Flash — 55 of 103 correct (53%)
4. Grok 4.6 — 55 of 103 correct (53%)
5. Qwen 3.8 Max — 55 of 103 correct (53%)
6. GPT-5.6 Luna — 53 of 103 correct (51%)
7. Kimi K3 — 52 of 103 correct (50%)
8. DeepSeek V4 Pro — 51 of 103 correct (50%)
9. Mistral Medium 3.5 — 46 of 103 correct (45%)
10. MiMo V2.5 Pro — 46 of 103 correct (45%)
11. Cohere Command R+ (08-2024) — 29 of 43 correct (67%)
12. Llama 3.1 70B Instruct — 21 of 43 correct (49%)

The spread from first to last among the main ten is 12 percentage points, but most of the field sits in a narrow band: eight of the ten landed between 50% and 53%. GLM 5.3 is the clear outlier on the upside, while Mistral Medium 3.5 and MiMo V2.5 Pro share the bottom at 45%.

Movers and fallers

Comparing this week with the all-time rates shows which models beat or fell short of their long-run form. GLM 5.3 over-performed most: its all-time rate is 47.6% (358 of 752), so a 57% week is a rise of roughly 9 percentage points. Kimi K3 also moved up, from an all-time 47.7% (359 of 752) to 50% this week. Gemini 3.8 Flash, Grok 4.6 and Qwen 3.8 Max each gained ground relative to their all-time rates of 48.6%, 48.2% and 47.8% respectively, all finishing on 53%.

Claude Sonnet 5, the all-time leader at 51.3% (675 of 1315), posted 53% — a modest over-performance. GPT-5.6 Luna edged above its 46.8% all-time rate with 51%. DeepSeek V4 Pro, second on the all-time list at 49.3% (648 of 1315), under-performed slightly at 50%... but that 50% is actually above its long-run rate, so the week was a small positive for DeepSeek rather than a fall. The genuine under-performers were Mistral Medium 3.5 (45% this week vs 47.4% all-time) and MiMo V2.5 Pro (45% vs 46.0%). Both dropped roughly 1–2 points below their historical averages.

Among the two extra systems, Cohere Command R+ (08-2024) recorded 67% on 43 matches, but with no all-time row in the dossier it cannot be compared on the same basis. Llama 3.1 70B Instruct finished at 49% on 43 matches. Their smaller match counts make weekly swings less meaningful.

All-time leaderboard

1. Claude Sonnet 5 — 675 of 1315 correct (51.3%)
2. DeepSeek V4 Pro — 648 of 1315 correct (49.3%)
3. Gemini 3.8 Flash — 638 of 1314 correct (48.6%)
4. Grok 4.6 — 634 of 1315 correct (48.2%)
5. Qwen 3.8 Max — 629 of 1315 correct (47.8%)
6. Mistral Medium 3.5 — 623 of 1315 correct (47.4%)
7. GPT-5.6 Luna — 616 of 1315 correct (46.8%)
8. MiMo V2.5 Pro — 605 of 1315 correct (46.0%)
9. Kimi K3 — 359 of 752 correct (47.7%)
10. GLM 5.3 — 358 of 752 correct (47.6%)

Claude Sonnet 5 retains the top spot all-time, 27 correct picks ahead of DeepSeek V4 Pro. Kimi K3 and GLM 5.3 have smaller samples, 752 decided matches each, so their positions are more sensitive to future results. The full list is available on the live AI leaderboard.

What it means for next week

Weekly results are a small sample. With 103 decided matches, a difference of a few correct picks can move a model several places. GLM 5.3's 57% is the standout number, but it covers only one week; the all-time table still places it ninth at 47.6%. The same caution applies to any single week's ordering.

Consensus reliability is a separate question. When several models agree on a pick, the group is often more stable than any individual model's weekly rate. This week, five models — Claude Sonnet 5, Gemini 3.8 Flash, Grok 4.6, Qwen 3.8 Max and GLM 5.3 — finished at or above 53%, suggesting broad agreement on many matches. That kind of clustering is worth watching across multiple weeks before drawing conclusions.

Next week's fixtures will add to the sample and test whether GLM 5.3's run continues or reverts toward its long-run rate. The live AI leaderboard updates as matches are decided.

Live table: the AI model leaderboard updates after every result, and the method page explains how picks are graded.