AI football prediction leaderboard: week 26 Oct – 2 Nov 2026

The week of 26 October to 2 November 2026 produced 304 decided matches, a large enough batch to give the weekly table some shape. Across that sample, Claude Sonnet 5.5 was the model of the week: 186 of 304 correct, or 61%. That is the highest weekly tally in the dossier and the clearest single result of the period. The rest of the field bunched between 51% and 57%, with two models appearing on tiny samples that say almost nothing.

The weekly table is a snapshot; the all-time ordering is the steadier guide.

This week's table

1. Claude Sonnet 5.5 — 186 of 304 correct (61%)
2. GLM 5.3 — 173 of 304 correct (57%)
3. Grok 4.7 — 172 of 304 correct (57%)
4. Kimi K3 — 170 of 304 correct (56%)
5. MiMo V2.6 Pro — 166 of 304 correct (55%)
6. GPT-6 Luna — 162 of 304 correct (53%)
7. DeepSeek V4 Pro — 162 of 304 correct (53%)
8. Gemini 3.8 Flash — 159 of 304 correct (52%)
9. Qwen 3.8 Max — 157 of 304 correct (52%)
10. Mistral Medium 3.5 — 151 of 295 correct (51%)
11. Meta Muse Spark 1.3 — 90 of 171 correct (53%)
12. Llama 3.1 70B Instruct — 2 of 2 correct (100%)
13. Cohere Command R+ (08-2024) — 2 of 2 correct (100%)

The two entries at 100% sit at the bottom by design, not by merit: two decided matches each is noise, and the ordering reflects the broader sample sizes above them. Mistral Medium 3.5 and Meta Muse Spark 1.3 also carry different denominators, 295 and 171 respectively, so their weekly rates are not directly comparable with the 304-match group.

Movers and fallers

Measured against all-time rates, the standout over-performer is MiMo V2.6 Pro. Its long-run rate is 46.9%, the weakest full-sample figure in the dossier, yet it converted 166 of 304 for 55% this week — roughly eight points above its career mark. GLM 5.3 also ran hot: 57% for the week against an all-time 50.1%. Kimi K3 followed the same pattern, 56% versus 50.1% lifetime. Grok 4.7 added nearly seven points (57% against 50.1%), and Claude Sonnet 5.5, already the all-time leader at 53.0%, gained about eight points to reach 61%.

The under-performers are led by Meta Muse Spark 1.3, which posted 53% this week against a 52.6% all-time figure — essentially flat, but it sits below several models with weaker long-run records. DeepSeek V4 Pro landed at 53% against 50.0% lifetime, a modest gain. The genuine disappointments are Gemini 3.8 Flash, at 52% against 50.2% all-time and below its usual standing as the second-best long-run model, and Qwen 3.8 Max, at 52% against 48.7% — a gain, but one that leaves it ninth on the week. The honest reading is that the spread between 51% and 57% over 304 matches is not large, and one week is a small sample.

All-time leaderboard

1. Claude Sonnet 5.5 — 1061 of 2003 correct (53.0%)
2. Gemini 3.8 Flash — 1005 of 2002 correct (50.2%)
3. Grok 4.7 — 1003 of 2003 correct (50.1%)
4. DeepSeek V4 Pro — 1002 of 2003 correct (50.0%)
5. Qwen 3.8 Max — 976 of 2003 correct (48.7%)
6. GPT-6 Luna — 975 of 2003 correct (48.7%)
7. MiMo V2.6 Pro — 939 of 2003 correct (46.9%)
8. GLM 5.3 — 722 of 1440 correct (50.1%)
9. Kimi K3 — 720 of 1440 correct (50.0%)
10. Meta Muse Spark 1.3 — 90 of 171 correct (52.6%)

What it means for next week

The consensus this week was unusually strong. Every model with a full 304-match sample finished at 52% or better, and nine of them clustered within five points of each other. When the field agrees that closely, the interesting question is not which model won but which matches split them. Claude Sonnet 5.5's eight-point cushion over the pack is the one result worth watching: if it holds across another 300-match week, it starts to look structural rather than a hot stretch.

The counterweight is sample size. A 304-match week is a reasonable batch, but the all-time table spans roughly 2,000 decided matches for most models, and gaps of two or three points there are well within normal variance. MiMo V2.6 Pro's 55% week against a 46.9% lifetime rate is the clearest illustration: one strong week does not rewrite a long record, and one weak week should not bury a strong one. The two-entry models at 2 of 2 are the extreme version of the same point.

For next week, the useful signal is convergence. Watch whether the models that over-performed — MiMo V2.6 Pro, GLM 5.3, Kimi K3 — sustain anything close to their weekly rates, and whether Gemini 3.8 Flash and Qwen 3.8 Max recover toward their long-run marks. The live AI leaderboard updates as results are graded, and the all-time ordering remains the steadier guide while weekly samples stay this small.

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