AI football prediction leaderboard: week 2 Nov – 9 Nov 2026

The week of 2 Nov – 9 Nov 2026 put 334 decided matches in front of the panel, and the results were tight enough that the top of the table was decided by a handful of correct picks. The model of the week is MiMo V2.6 Pro, which called 177 of 334 matches correctly (53%), finishing one pick clear of Claude Sonnet 5.5 on 174 of 334 (52%). A further model, Mistral Medium 3.5, appears in the weekly table with 154 of 307 correct (50%) but has no all-time entry, so it is ranked inside the week only. Across the ten long-running models the spread from first to last is four percentage points, which is the kind of gap a single week can produce without any underlying change in skill.

Ten forecasting systems, one week of 334 decided matches.

This week's table

Ranking for the 2 Nov – 9 Nov 2026 window, ordered by correct picks. Percentages are as reported.

1. MiMo V2.6 Pro — 177 of 334 correct (53%)
2. Claude Sonnet 5.5 — 174 of 334 correct (52%)
3. GPT-6 Luna — 169 of 334 correct (51%)
4. Meta Muse Spark 1.3 — 169 of 334 correct (51%)
5. Qwen 3.8 Max — 167 of 334 correct (50%)
6. Grok 4.7 — 165 of 334 correct (49%)
7. DeepSeek V4 Pro — 165 of 334 correct (49%)
8. Kimi K3 — 165 of 334 correct (49%)
9. Gemini 3.8 Flash — 164 of 334 correct (49%)
10. GLM 5.3 — 164 of 334 correct (49%)
11. Mistral Medium 3.5 — 154 of 307 correct (50%)

Movers and fallers

The clearest over-performer is MiMo V2.6 Pro. Its all-time rate is 47.7%, the lowest of the ten established models, yet it led the week at 53%. That is a five-point jump above its long-run number and the reason it takes the weekly title. GPT-6 Luna also ran ahead of itself, posting 51% against an all-time 48.9%, and Qwen 3.8 Max did the same at 50% against 48.9%. Meta Muse Spark 1.3 matched GPT-6 Luna on 51% for the week, a touch below its 51.3% all-time mark, so it held roughly to form rather than gaining.

At the other end, Gemini 3.8 Flash was the notable faller. It sits second on the all-time list at 50.0% but managed 164 of 334 (49%) this week, a small drop that still cost it places in a compressed field. Claude Sonnet 5.5, the all-time leader at 52.8%, came in at 52% — close to its long-run rate and enough for second on the week. Grok 4.7 (49% against 50.0%), DeepSeek V4 Pro (49% against 49.9%), Kimi K3 (49% against 49.9%) and GLM 5.3 (49% against 49.9%) all landed within a point of their all-time figures. The honest reading is that most of these moves are noise: with 334 matches, one or two extra correct calls separate mid-table from the top, and a single week is a small sample against careers of 1,770 to 2,333 decided matches.

All-time leaderboard

Winner picks on decided matches, as published.

1. Claude Sonnet 5.5 — 1232 of 2333 correct (52.8%)
2. Gemini 3.8 Flash — 1167 of 2332 correct (50.0%)
3. Grok 4.7 — 1166 of 2333 correct (50.0%)
4. DeepSeek V4 Pro — 1165 of 2333 correct (49.9%)
5. GPT-6 Luna — 1142 of 2333 correct (48.9%)
6. Qwen 3.8 Max — 1141 of 2333 correct (48.9%)
7. MiMo V2.6 Pro — 1114 of 2333 correct (47.7%)
8. GLM 5.3 — 884 of 1770 correct (49.9%)
9. Kimi K3 — 883 of 1770 correct (49.9%)
10. Meta Muse Spark 1.3 — 257 of 501 correct (51.3%)

What it means for next week

The weekly table and the all-time table disagree, and that disagreement is the useful signal. MiMo V2.6 Pro won the week but remains seventh all-time at 47.7%, below the 50% line that separates a coin flip from a genuine edge. Claude Sonnet 5.5 had a quieter week at 52% but still holds first place at 52.8% over 2,333 decided matches. When a model's weekly result is five points above its career rate, the likelier explanation is variance rather than a change in form, and the next window will test whether MiMo can repeat 53% or falls back toward its 47.7% baseline.

Consensus reliability is the more practical question. Nine of the eleven entries this week fell between 49% and 51%, and the all-time list is similarly bunched between 47.7% and 52.8% for the models with full histories. That means a single model's pick carries modest information on any one match; agreement across several models is the stronger indicator, and the leaderboard is built to show exactly where that agreement concentrates.

Watch the sample sizes as much as the ranks. Meta Muse Spark 1.3 has 501 decided matches, GLM 5.3 and Kimi K3 have 1,770, and the leaders have more than 2,300. Rates from the shorter records will move faster and should be read with that in mind. The live AI leaderboard updates as matches are decided, and next week's window will show whether MiMo's 53% was a spike or the start of a climb.

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