AI football prediction leaderboard: week 12 Oct – 19 Oct 2026

Week 12 Oct – 19 Oct 2026: Gemini 3.8 Flash takes the week

The week from 12 October to 19 October 2026 produced 117 decided matches, enough to rank ten of the twelve tracked systems on a single week's output. Twelve models appear in the last-seven-days table, but two of them — Llama 3.1 70B Instruct and Cohere Command R+ (08-2024) — logged only 48 decided matches each, so their week is not directly comparable to the others.

Gemini 3.8 Flash is the model of the week, with 64 of 117 correct (55%). That is the only tally above 54% in the cohort, and it came against a long-run rate of 49.1%, which is the second-best all-time figure on the board. Claude Sonnet 5 finished a single match behind on 63 of 117 correct (54%), and Mistral Medium 3.5 and Grok 4.6 tied on 61 of 117 correct (52%).

The spread across the week is wide at the bottom. MiMo V2.5 Pro converted 44 of 117 correct (38%), nine percentage points below the next-lowest model and well under its own all-time rate of 45.4%. That single week does more to define the table than any of the mid-pack results.

This week's table

1. Gemini 3.8 Flash — 64 of 117 correct (55%)
2. Claude Sonnet 5 — 63 of 117 correct (54%)
3. Mistral Medium 3.5 — 61 of 117 correct (52%)
4. Grok 4.6 — 61 of 117 correct (52%)
5. DeepSeek V4 Pro — 60 of 117 correct (51%)
6. Qwen 3.8 Max — 59 of 117 correct (50%)
7. GPT-5.6 Luna — 58 of 117 correct (50%)
8. GLM 5.3 — 57 of 117 correct (49%)
9. Kimi K3 — 55 of 117 correct (47%)
10. MiMo V2.5 Pro — 44 of 117 correct (38%)
11. Llama 3.1 70B Instruct — 24 of 48 correct (50%)
12. Cohere Command R+ (08-2024) — 24 of 48 correct (50%)

The order is set by raw correct picks, not by percentage, which is why the two 48-match models sit at the bottom despite a 50% rate that would place them mid-table. The 50% mark is not a high bar: seven of the twelve systems finished at or above it, and four of those were within one match of each other.

Movers and fallers

Comparing the week against the all-time column separates genuine movement from noise. Gemini 3.8 Flash gained the most ground: 55% for the week against 49.1% all-time, a rise of roughly six percentage points. Claude Sonnet 5 also ran hot at 54% against a 51.6% long-run rate, and Mistral Medium 3.5 moved up from 47.7% to 52%.

Three models beat their long-run figures by smaller margins. Grok 4.6 landed at 52% against 48.6%, DeepSeek V4 Pro at 51% against 49.5%, and Qwen 3.8 Max at 50% against 48.1%. GPT-5.6 Luna came in at 50% against 47.1%, a modest improvement on its all-time position of seventh.

The under-performers are more instructive. GLM 5.3 posted 49% for the week against an all-time 47.8%, which is close to flat, but Kimi K3 slipped to 47% from 47.7% and MiMo V2.5 Pro fell sharply to 38% from 45.4% — the largest negative gap in the cohort. Llama 3.1 70B Instruct and Cohere Command R+ (08-2024) recorded 50% on 48 matches each; neither appears in the all-time list in this dossier, so no long-run comparison is available for them.

It is worth stating plainly that a 117-match week is a small sample. A swing of six percentage points on 117 matches is roughly seven matches. That is inside the range a coin-flip-heavy competition can produce by chance alone, and the all-time rates are the more stable signal.

All-time leaderboard

1. Claude Sonnet 5 — 738 of 1431 correct (51.6%)
2. DeepSeek V4 Pro — 708 of 1431 correct (49.5%)
3. Gemini 3.8 Flash — 702 of 1430 correct (49.1%)
4. Grok 4.6 — 695 of 1431 correct (48.6%)
5. Qwen 3.8 Max — 688 of 1431 correct (48.1%)
6. Mistral Medium 3.5 — 683 of 1431 correct (47.7%)
7. GPT-5.6 Luna — 674 of 1431 correct (47.1%)
8. MiMo V2.5 Pro — 649 of 1431 correct (45.4%)
9. GLM 5.3 — 415 of 868 correct (47.8%)
10. Kimi K3 — 414 of 868 correct (47.7%)

The ranking is by total correct picks, which is why GLM 5.3 and Kimi K3 sit ninth and tenth despite percentage rates that would put them above several models with more matches. The gap from first to eighth is 6.2 percentage points; the gap from first to tenth, on the available rates, is under four.

What it means for next week

Consensus reliability is the practical question. When the top six models in a given week all land between 50% and 55%, the aggregate is a narrow band, and the differences between individual systems are smaller than the week-to-week variation. The useful read is not which model led a single week but whether a model consistently sits above its own baseline. On that measure, Gemini 3.8 Flash's week is the strongest result, because it combined the highest tally with a clear gain over its all-time rate.

The models to watch are the ones whose week diverged most from their history. MiMo V2.5 Pro's 38% is the outlier to track: a second week near that level would suggest something more than variance, while a return toward 45% would confirm the week as an anomaly. GLM 5.3 and Kimi K3 are effectively level with their long-run rates, so neither has yet shown a trend in either direction. Llama 3.1 70B Instruct and Cohere Command R+ (08-2024) need more decided matches before any comparison is meaningful.

Sample size remains the constraint on every claim here. The full picture, updated after each matchday, is on the live AI leaderboard, where the all-time columns carry far more weight than any single seven-day window. Next week adds another set of decided matches; until then, the 12–19 October table should be read as one data point, not a ranking shift.

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