AI football prediction leaderboard: week 19 Oct – 26 Oct 2026

The week of 19 Oct – 26 Oct 2026 produced 274 decided matches on the TuringStats board, plus 51 matches settled by two models newly added to the week's sample. The model of the week is Gemini 3.8 Flash, which called 147 of 274 correctly (54%). That is the best tally of any model with a full week of picks and the only one to clear the 52% mark. Claude Sonnet 5, the all-time leader, finished third on 140 of 274 (51%), a solid but unremarkable week by its own standards. The spread between first and last among the ten full-week models was nine percentage points, which is roughly the scale of noise a 274-match sample can produce.

Ten forecasting systems, one week of 274 decided matches.

This week's table

Ranking of every model that made picks in the week, ordered by correct calls. Two models appear with a 51-match sample and are not directly comparable to the ten full-week entries.

1. Gemini 3.8 Flash — 147 of 274 correct (54%)
2. GPT-5.6 Luna — 142 of 274 correct (52%)
3. Claude Sonnet 5 — 140 of 274 correct (51%)
4. Grok 4.6 — 139 of 274 correct (51%)
5. Kimi K3 — 139 of 274 correct (51%)
6. GLM 5.3 — 137 of 274 correct (50%)
7. DeepSeek V4 Pro — 134 of 274 correct (49%)
8. Qwen 3.8 Max — 134 of 274 correct (49%)
9. MiMo V2.6 Pro — 127 of 274 correct (46%)
10. Mistral Medium 3.5 — 122 of 274 correct (45%)
11. Llama 3.1 70B Instruct — 29 of 51 correct (57%)
12. Cohere Command R+ (08-2024) — 24 of 51 correct (47%)

Movers and fallers

Comparing the week against all-time rates separates genuine form from sample noise. Gemini 3.8 Flash over-performed its long-run 49.8% by more than four points, the largest positive gap in the table, and that gap is what earns it the week. GPT-5.6 Luna also beat its all-time 47.9% by roughly four points, landing at 52%. Grok 4.6 ran about two points above its 48.9% career rate, and Kimi K3 added roughly three points to its 48.4% baseline.

Claude Sonnet 5 came in at 51%, marginally below its all-time 51.5%, a week that looks flat rather than poor. DeepSeek V4 Pro was close to its 49.4% norm at 49%. Qwen 3.8 Max sat a fraction above its 48.2% career figure.

The fallers are more pronounced. Mistral Medium 3.5 dropped to 45% against an all-time 47.2%, a shortfall of more than two points and the weakest full-week tally on the board. MiMo V2.6 Pro, the lowest all-time performer at 45.5%, came in at 46% — essentially on trend. GLM 5.3 finished at 50%, about two points above its 48.3% baseline, though its career sample of 1142 matches is smaller than the leaders'. The two short-sample entries, Llama 3.1 70B Instruct at 57% and Cohere Command R+ (08-2024) at 47%, rest on 51 matches each and should not be read as form until the sample grows. A week is a small sample for every model here; the all-time column remains the better guide.

All-time leaderboard

Winner picks on decided matches, as published on the board.

1. Claude Sonnet 5 — 878 of 1705 correct (51.5%)
2. Gemini 3.8 Flash — 849 of 1704 correct (49.8%)
3. DeepSeek V4 Pro — 842 of 1705 correct (49.4%)
4. Grok 4.6 — 834 of 1705 correct (48.9%)
5. Qwen 3.8 Max — 822 of 1705 correct (48.2%)
6. GPT-5.6 Luna — 816 of 1705 correct (47.9%)
7. Mistral Medium 3.5 — 805 of 1705 correct (47.2%)
8. MiMo V2.6 Pro — 776 of 1705 correct (45.5%)
9. Kimi K3 — 553 of 1142 correct (48.4%)
10. GLM 5.3 — 552 of 1142 correct (48.3%)

What it means for next week

The useful signal this week is not who won but how tightly the field clustered. Six of the ten full-week models landed between 49% and 54%, and the ordering shifted by several places relative to the all-time table. Gemini 3.8 Flash's week-leading 54% is a four-point jump on a career rate of 49.8%; at 274 matches, a swing of that size is well within what chance alone can produce. Reading it as a durable improvement would be premature. The same caution applies downward: Mistral Medium 3.5's 45% is its worst weekly showing in this sample, but its all-time 47.2% is only about two points better, so the week says more about variance than about a collapse.

Consensus reliability follows the same logic. Where models agree on a winner, the group has historically been harder to beat than any single entrant, but agreement is not a guarantee, and this week's spread shows how easily a short window can reorder the table. The gap between the top all-time model, Claude Sonnet 5 at 51.5%, and the bottom, MiMo V2.6 Pro at 45.5%, is six points across more than 1700 matches each — a real but narrow margin. Anyone treating a single week's ranking as a stable forecast is reading noise.

The practical takeaway for next week is to weight the all-time record more heavily than the last seven days, and to treat the two 51-match entries as provisional until their samples grow. The full rankings, updated after every decided match, are on the live AI leaderboard.

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