Every AI Model Lost Money Betting on Soccer — The KellyBench Study Explained

Last Updated on October 14, 2026

Updated October 2026: rewritten around the full KellyBench paper (five runs per model) and General Reasoning’s latest leaderboard, which now covers 12 models. Every one still loses money.

AI can write code, pass exams and argue about football tactics all day.

So can it beat the bookies? A careful test says no.

Not one top AI model made money.

What I find most interesting is why they lost. They knew plenty about football.

They just couldn’t turn that knowledge into good bets.

Can AI beat sports betting?

Not yet, on this evidence. In the KellyBench test, no AI model made money on average across a season.

A few single runs did end in profit. But the same models lost money on most other runs.

Betting is hard for everyone. Bookmakers build their margin into the odds, so most bettors lose over time.

To win, you need to spot when the odds are wrong, and bet the right amount when they are. That’s exactly where the AI models struggled.

If gambling is causing you problems, help is free. The NHS guide to gambling addiction lists support, including the National Gambling Helpline on 0808 8020 133 in the UK.

In the US, the National Council on Problem Gambling runs a free helpline at 1-800-GAMBLER.

What KellyBench tested

KellyBench comes from General Reasoning, an AI company. Its paper was posted on arXiv in April 2026.

The test replays the 2023-24 English Premier League season, one matchday at a time. Each AI starts with £100,000 and tries to grow it.

The models get real bookmaker odds, team statistics, line-ups and years of past results. They have to build their own prediction models, decide which bets have an edge, and choose how much to stake.

Each run is long. The models used 500 to 1,000 tool calls per season.

The name comes from the Kelly criterion, a classic formula for bet size. It tells you to bet more when your edge is bigger, and nothing when you have no edge.

That sounds sensible. But if your odds are even a little wrong, Kelly makes you bet too much, and losses pile up fast.

What the research shows

Every model lost money

The paper tested five models, five times each, across the whole season.

Only 3 of the 25 runs made a profit. Each of those came from a model that lost money on average.

Gemini 3.1 Pro shows why one good run means little. It made 33.7% once, and lost everything in another run.

Three more models were tried in an earlier version of the paper and dropped. Grok 4.20 and Arcee Trinity mostly couldn’t even finish the season properly.

The newer models still lose

General Reasoning keeps a live leaderboard. By its latest update, it covered 12 models.

Claude Opus 4.7 now leads, with an average return of -3.7%. GPT-5.4 and GPT-5.5 come next, both losing money too.

So the newest models are closer to breaking even. None has crossed the line.

A simple old model beat most of them

The team also tested human-made strategies on the same season.

A classic football model from 1997, called Dixon-Coles, lost 15.4%. That still beat three of the five AI models in the paper.

A human with two years of experience building football models made 5.1% in a week of work. They placed just 39 bets, far fewer than the AI models.

An AI researcher with no betting background lost 4.3%. That’s still better than every model in the paper.

Why the AI models lost

The main problem was prediction. The models’ chances for each match were less accurate than the odds the bookmakers already offered.

When your odds are worse than the market’s, every “value bet” you find is usually a mistake.

The paper lists the most common failures across the 25 runs:

• In 22 runs, models had no general way to handle newly promoted teams.
• In 22 runs, models misjudged draws or long shots again and again.
• In 9 runs, models bet without any sound rule for stake size. In 7 more, they wrote Kelly code but never used it.
• In 8 runs, models declared the season “complete” while matches were still being played.
• In 6 runs, models lost the whole bankroll.

One Kimi K2.5 run bet 98% of its remaining money on a single game, Burnley against Luton, by accident.

Promoted teams were a big trap. The models had little recent data on those clubs, and few adjusted.

The researchers also rated each strategy against a 52-point checklist written with betting experts.

The best AI, Claude Opus 4.6, scored 26.5%. The human expert scored 73.1%.

What earlier research found

AI forecasters trail human experts

A 2024 study called ForecastBench tested AI models on 200 questions about future events. It compared them with expert human forecasters and the public.

Expert forecasters beat the best AI model, and the gap was clearly real. Like KellyBench, it found AI does worse here than on most other tests.

But AI can get close with help

A 2024 study from the University of California, Berkeley, built an AI system to answer forecasting questions. It searched for news and combined several predictions.

On average, it came close to a crowd of skilled human forecasters, and sometimes beat it. That suggests the right setup matters a lot.

How it fits together

AI models can reason about the future reasonably well. But turning those guesses into money, against a sharp market, is a much harder job.

KellyBench tests that harder job, over months, with real consequences for every mistake.

How much should you trust this?

Promising. The test is careful and repeated, but narrow.

What makes it convincing

• It uses real odds from a real season, so the market is a tough opponent.
• Each model ran five times, which shows luck from skill.
• The team checked human baselines, so we know the test can be beaten.
• The paper shares its failures openly, including runs where models broke the rules.

What makes me cautious

• It covers one league and one season.
• Each model worked alone, with one setup. Teams of AI agents weren’t tested.
• General Reasoning built the test and sells AI evaluation tools.
• Models change fast. The leaderboard has already moved since the paper.

What this means for you

• Don’t pay for “AI betting picks”. In the best test so far, the top models lost money.
• Watch for cherry-picked wins. One profitable run means little. Gemini made 33.7% once and lost everything in another run.
• Use AI for research, not decisions. Models are good at summarising news. They’re weak at pricing risk over time.
• Treat any betting as a cost. Set a limit you can afford to lose.

This explainer from Mutual Information shows how the Kelly criterion works, and why getting your odds slightly wrong is so costly:

If you’re curious how far AI agents have come in other tasks, we looked at what AI browser agents can really do.

What we still don’t know

1. Would teams of AI agents do better? The paper only tested single agents.
2. Does it hold in other sports? Only English football was tested.
3. How close are the newest models? The leaderboard keeps changing.
4. Is the market just too efficient? Even the human expert made only a small profit.

My take: knowing football isn’t the same as beating the bookies

What I like about KellyBench is that it measures something real. Not a quiz, but months of decisions where every mistake costs money.

The models clearly knew a lot about football. They still lost, because their numbers weren’t better than the market’s.

I’m also struck by how close the newest models are getting. That’s worth watching.

For now, though, the lesson is simple. If someone sells you AI betting tips, the best evidence says the AI would lose too.