AI football prediction leaderboard: 21 Sep – 28 Sep 2026

Week 21 Sep – 28 Sep 2026 saw 102 decided matches graded across the ten frontier models tracked by TuringStats. The model of the week is MiMo V2.5 Pro, which called 58 of 102 correctly (57%). That is the highest raw tally of the week and the highest hit rate, edging out Claude Sonnet 5 on 56 of 102 (55%). The spread from top to bottom was ten percentage points, which in a week of this size is a reminder that short windows move more than long-run skill does.

Ten models, 102 decided matches: the weekly table is a snapshot of form, not a revision of the all-time ranking.

This week's table

1. MiMo V2.5 Pro — 58 of 102 correct (57%)
2. Claude Sonnet 5 — 56 of 102 correct (55%)
3. Kimi K3 — 55 of 102 correct (54%)
4. GPT-5.6 Luna — 54 of 102 correct (53%)
5. Gemini 3.8 Flash — 54 of 102 correct (53%)
6. GLM 5.3 — 54 of 102 correct (53%)
7. Qwen 3.8 Max — 52 of 102 correct (51%)
8. DeepSeek V4 Pro — 51 of 102 correct (50%)
9. Grok 4.6 — 51 of 102 correct (50%)
10. Mistral Medium 3.5 — 48 of 102 correct (47%)

Movers and fallers

Comparing the last seven days with the all-time rates shows who ran hot and who ran cold. MiMo V2.5 Pro was the clearest over-performer: 57% this week against an all-time 46.0% (501 of 1088). That is an 11-point jump, and it is the kind of gap that a 102-match sample can produce without any underlying change in model quality. Claude Sonnet 5 also beat its long-run mark, 55% against 51.4% (559 of 1088), though by a smaller margin and from a higher base.

Kimi K3 and GLM 5.3 both cleared their all-time rates, 54% against 47.0% (247 of 525) and 53% against 45.3% (238 of 525) respectively. Those two have smaller all-time samples than the rest of the field, so their weekly swings carry less weight. GPT-5.6 Luna (53% vs 45.8%, 498 of 1088) and Gemini 3.8 Flash (53% vs 47.5%, 516 of 1087) also finished above their long-run lines.

On the other side, Mistral Medium 3.5 was the week's under-performer: 47% against an all-time 48.1% (523 of 1088). DeepSeek V4 Pro landed at 50% against 49.4% (538 of 1088), effectively flat. Grok 4.6 also came in at 50% against 48.1% (523 of 1088), a touch above its baseline. Qwen 3.8 Max finished at 51% against 47.3% (515 of 1088). The net picture is a week where most of the field beat its own long-run rate, which is itself a signal that the 102-match slate was not especially punishing.

All-time leaderboard

1. Claude Sonnet 5 — 559 of 1088 correct (51.4%)
2. DeepSeek V4 Pro — 538 of 1088 correct (49.4%)
3. Grok 4.6 — 523 of 1088 correct (48.1%)
4. Mistral Medium 3.5 — 523 of 1088 correct (48.1%)
5. Gemini 3.8 Flash — 516 of 1087 correct (47.5%)
6. Qwen 3.8 Max — 515 of 1088 correct (47.3%)
7. MiMo V2.5 Pro — 501 of 1088 correct (46.0%)
8. GPT-5.6 Luna — 498 of 1088 correct (45.8%)
9. Kimi K3 — 247 of 525 correct (47.0%)
10. GLM 5.3 — 238 of 525 correct (45.3%)

What it means for next week

One week of 102 matches is a small sample. The gap between MiMo V2.5 Pro's 57% and its 46.0% all-time rate is large enough to look meaningful, but it is well within the range that random variation produces over a single slate. The same applies at the other end: Mistral Medium 3.5's 47% week is not evidence that the model has regressed from its 48.1% baseline. Treat the weekly table as a snapshot of form, not a revision of the long-run ranking.

The more stable signal is the all-time list, where Claude Sonnet 5 leads at 51.4% (559 of 1088) and the top eight models sit within roughly six points of each other. That clustering matters for anyone reading the live AI leaderboard: when models are this close, a single week can reshuffle the order without any model actually changing. The two entries with 525-match samples, Kimi K3 and GLM 5.3, will need more decided matches before their all-time rates carry the same weight as the rest.

For next week, the practical takeaway is to watch whether MiMo V2.5 Pro holds anywhere near 57% and whether Claude Sonnet 5 continues to sit above its 51.4% baseline. If the top group stays bunched, the weekly winner will keep rotating. The live AI leaderboard updates as results are graded, and it remains the best place to separate a genuine run from a small-sample spike.

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