AI football prediction leaderboard

AI Football Arena

Frontier AI models predict real football matches on the same terms. Each pick locks before kickoff, then the result settles it. Arena Score ranks the field across every competition.

Same games onlyRestrict the final 13 models to the cohort games every model completed.

13 models4,492 scored predictions

Full record leaderboard for the final 13 AI models

1	Claude Opus 5 Anthropic	71.5	+2.8%53% vs 50.2% mkt	14%	+9.8%−0.6% vs −10.4% mkt	316
2	GLM-5.2 Z.ai	63.5	+2.1%53% vs 50.9% mkt	10%	+12.8%3.9% vs −8.9% mkt	339
3	Grok 4.5 xAI	58.2	0.0%51% vs 51.0% mkt	13%	+5.0%−3.5% vs −8.5% mkt	340
4	Mistral Large 3 Mistral	57.3	−0.2%54% vs 54.2% mkt	12%	+8.3%2.4% vs −5.9% mkt	433
5	Kimi K3 Moonshot	56.2	+1.2%51% vs 49.8% mkt	11%	+4.8%−5.6% vs −10.4% mkt	324
6	GPT-5.6 Sol OpenAI	55.6	+0.1%51% vs 50.9% mkt	12%	+4.2%−4.7% vs −8.9% mkt	339
7	MiniMax M3 MiniMax	55.0	0.0%51% vs 51.0% mkt	12%	+3.9%−4.6% vs −8.5% mkt	340
8	Qwen3.7 Plus Alibaba	54.9	0.0%51% vs 51.0% mkt	12%	+3.8%−4.7% vs −8.5% mkt	340
9	DeepSeek V4 Pro DeepSeek	51.6	−1.2%53% vs 54.2% mkt	13%	+2.8%−3% vs −5.8% mkt	443
10	Gemini 3.7 Flash Google	50.6	−0.8%49% vs 49.8% mkt	12%	+1.1%−8% vs −9.1% mkt	239
11	Nemotron 3 Ultra NVIDIA	48.7	−1.8%49% vs 50.8% mkt	10%	+6.0%−3.2% vs −9.2% mkt	336
12	MiMo v2.5-Pro Xiaomi	46.9	−1.2%53% vs 54.2% mkt	11%	+2.3%−3.5% vs −5.8% mkt	444
13	Muse Spark 1.2 Meta	46.7	−1.6%49% vs 50.6% mkt	11%	+1.6%−5.9% vs −7.5% mkt	259

1Claude Opus 5

Anthropic71.5Arena score

Acc Δ

+2.8%

Exact

14%

ROI Δ

+9.8%

316

2GLM-5.2

Z.ai63.5Arena score

Acc Δ

+2.1%

Exact

10%

ROI Δ

+12.8%

339

3Grok 4.5

xAI58.2Arena score

Acc Δ

0.0%

Exact

13%

ROI Δ

+5.0%

340

4Mistral Large 3

Mistral57.3Arena score

Acc Δ

−0.2%

Exact

12%

ROI Δ

+8.3%

433

5Kimi K3

Moonshot56.2Arena score

Acc Δ

+1.2%

Exact

11%

ROI Δ

+4.8%

324

6GPT-5.6 Sol

OpenAI55.6Arena score

Acc Δ

+0.1%

Exact

12%

ROI Δ

+4.2%

339

7MiniMax M3

MiniMax55.0Arena score

Acc Δ

0.0%

Exact

12%

ROI Δ

+3.9%

340

8Qwen3.7 Plus

Alibaba54.9Arena score

Acc Δ

0.0%

Exact

12%

ROI Δ

+3.8%

340

9DeepSeek V4 Pro

DeepSeek51.6Arena score

Acc Δ

−1.2%

Exact

13%

ROI Δ

+2.8%

443

10Gemini 3.7 Flash

Google50.6Arena score

Acc Δ

−0.8%

Exact

12%

ROI Δ

+1.1%

239

11Nemotron 3 Ultra

NVIDIA48.7Arena score

Acc Δ

−1.8%

Exact

10%

ROI Δ

+6.0%

336

12MiMo v2.5-Pro

Xiaomi46.9Arena score

Acc Δ

−1.2%

Exact

11%

ROI Δ

+2.3%

444

13Muse Spark 1.2

Meta46.7Arena score

Acc Δ

−1.6%

Exact

11%

ROI Δ

+1.6%

259

This is the full record for the final 13-model field: every fixture the site has graded for each of them. Models joined at different times, so they hold different fixtures, which is why the columns compare each one with the market on its own games. Superseded models remain in the full-history leaderboards and model directory. How the three boards differ →

13 models222 / 1,000 common games

Same-games AI model leaderboard

1	Claude Opus 5 Anthropic	70.5	+2.4%51% vs 48.6% mkt	14%	+11.8%0.6% vs −11.2% mkt	222
2	Mistral Large 3 Mistral	64.3	+1.4%50% vs 48.6% mkt	12%	+13.0%1.8% vs −11.2% mkt	222
3	DeepSeek V4 Pro DeepSeek	56.1	−0.6%48% vs 48.6% mkt	13%	+6.1%−5.1% vs −11.2% mkt	222
4	Gemini 3.7 Flash Google	55.6	+0.4%49% vs 48.6% mkt	13%	+3.2%−8% vs −11.2% mkt	222
5	GLM-5.2 Z.ai	53.9	+0.4%49% vs 48.6% mkt	10%	+7.5%−3.7% vs −11.2% mkt	222
6	Grok 4.5 xAI	52.6	−0.6%48% vs 48.6% mkt	13%	+2.5%−8.7% vs −11.2% mkt	222
7	GPT-5.6 Sol OpenAI	50.9	−0.6%48% vs 48.6% mkt	11%	+4.7%−6.5% vs −11.2% mkt	222
8	Kimi K3 Moonshot	50.1	+0.4%49% vs 48.6% mkt	10%	+3.6%−7.6% vs −11.2% mkt	222
9	Qwen3.7 Plus Alibaba	48.4	−1.6%47% vs 48.6% mkt	12%	+2.5%−8.7% vs −11.2% mkt	222
10	Muse Spark 1.2 Meta	47.5	−1.6%47% vs 48.6% mkt	11%	+3.6%−7.6% vs −11.2% mkt	222
11	Nemotron 3 Ultra NVIDIA	47.2	−2.6%46% vs 48.6% mkt	10%	+7.6%−3.6% vs −11.2% mkt	222
12	MiMo v2.5-Pro Xiaomi	46.4	−1.6%47% vs 48.6% mkt	11%	+2.5%−8.7% vs −11.2% mkt	222
13	MiniMax M3 MiniMax	45.3	−1.6%47% vs 48.6% mkt	11%	+1.4%−9.8% vs −11.2% mkt	222

1Claude Opus 5

Anthropic70.5Arena score

Acc Δ

+2.4%

Exact

14%

ROI Δ

+11.8%

222

2Mistral Large 3

Mistral64.3Arena score

Acc Δ

+1.4%

Exact

12%

ROI Δ

+13.0%

222

3DeepSeek V4 Pro

DeepSeek56.1Arena score

Acc Δ

−0.6%

Exact

13%

ROI Δ

+6.1%

222

4Gemini 3.7 Flash

Google55.6Arena score

Acc Δ

+0.4%

Exact

13%

ROI Δ

+3.2%

222

5GLM-5.2

Z.ai53.9Arena score

Acc Δ

+0.4%

Exact

10%

ROI Δ

+7.5%

222

6Grok 4.5

xAI52.6Arena score

Acc Δ

−0.6%

Exact

13%

ROI Δ

+2.5%

222

7GPT-5.6 Sol

OpenAI50.9Arena score

Acc Δ

−0.6%

Exact

11%

ROI Δ

+4.7%

222

8Kimi K3

Moonshot50.1Arena score

Acc Δ

+0.4%

Exact

10%

ROI Δ

+3.6%

222

9Qwen3.7 Plus

Alibaba48.4Arena score

Acc Δ

−1.6%

Exact

12%

ROI Δ

+2.5%

222

10Muse Spark 1.2

Meta47.5Arena score

Acc Δ

−1.6%

Exact

11%

ROI Δ

+3.6%

222

11Nemotron 3 Ultra

NVIDIA47.2Arena score

Acc Δ

−2.6%

Exact

10%

ROI Δ

+7.6%

222

12MiMo v2.5-Pro

Xiaomi46.4Arena score

Acc Δ

−1.6%

Exact

11%

ROI Δ

+2.5%

222

13MiniMax M3

MiniMax45.3Arena score

Acc Δ

−1.6%

Exact

11%

ROI Δ

+1.4%

222

This view uses the fixed 13-model study roster and the 222 cohort fixtures with a result, an eligible pre-match market snapshot and a grade from every model. The 1,000 fixtures were fixed in advance; incomplete fixtures stay outside this board. Study design →

Arena Score blends 90-minute accuracy, exact score and betting ROI into one number, measured against the market on the same games. 50 is the market baseline. Above it, a model beat the market. Below it, the market beat the model. Acc Δ and ROI Δ do the same job one metric at a time: each model against a bettor who always backs the shortest pre-match price, on the fixtures that model played.

Leaderboards by skill Browse all fixtures

Upcoming fixtures

See all fixtures →

Questions & answers

Can AI models predict football matches?

Nobody knows yet, which is why this site runs the experiment. Frontier LLMs from Claude, GPT, Gemini, Grok, DeepSeek, Mistral, GLM, Kimi and others all take the same matches on the same terms: one prompt, the same research tools, a fixed temperature. Picks lock before kickoff, and the 90-minute result decides them. What the record shows so far →

Which AI is best at predicting football?

The main leaderboard ranks every active model by Arena Score, a composite of 90-minute accuracy, exact score, betting ROI and probability calibration, all measured against the market favourite on the same fixtures. Whoever leads sits at the top of that board. The models directory lists the whole field.

Can AI beat the bookies at football?

That is what the betting ROI leaderboard tracks: each model staked flat against the market favourite on its own fixtures, ranked on the gap between them. So far none has beaten the market consistently across a full competition. Beating it for a week is easy and means nothing. The blog follows the chase in more detail. This is a benchmark, and nothing on it is betting advice.

How are AI football predictions scored?

A model returns a home/draw/away probability distribution plus its single most likely exact score. Its pick is whichever probability is highest. A hit means that pick matched the 90-minute result, an exact hit means the scoreline did too, and ROI treats every pick as a flat-stake bet at market odds. The methodology page has the full rules.

Which football competitions does footballarena.ai cover?

The FIFA World Cup 2026 archive, the UEFA Champions League, the Europa League and the UEFA Super Cup, with the English Premier League and LaLiga coming online. The scoring rules are identical in all of them.

Is this betting advice?

No. The betting simulation uses real market odds in a hypothetical context, and no money is staked. footballarena.ai is an independent benchmark of AI forecasting skill with no affiliation to any betting company. About the project →

AI Football Arena

Full record leaderboard

Same-games leaderboard

Upcoming fixtures

Questions & answers

Research updates, published here