Can AI Beat Sportsbooks? The 2026 World Cup Gave Us an Awkward Answer

The 2026 World Cup offered AI researchers something rare: 104 major matches that happened after the models' training cutoffs, with public team news, live odds and unambiguous results.

Four frontier models were sent to research each fixture, produce probabilities and place a virtual $100 bet. The bookmaker market stood beside them as the fifth competitor.

The machines sounded intelligent. The market was harder to beat.

Accuracy was not the same as profit

The WC2026-Agents benchmark compared Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro and Grok with pre-match 1X2 betting odds. The paper reports that none of the four models beat the market's Brier score, a measure of probabilistic accuracy. A naive strategy of backing the market favourite also out-earned all four agents. [1]

That does not mean every model lost money. Reported betting returns ranged from roughly -18% to +10%, showing that forecasting quality, price selection and staking decisions can produce different financial outcomes even when top picks look similar.

The models selected the same most likely outcome in 92% of matches. Their polished explanations concealed a striking amount of consensus.

Sportsbooks already use much of the obvious information

An AI can search injuries, form, expected line-ups, travel and tactical commentary. So can traders, data suppliers and thousands of market participants.

The question is not whether the model finds relevant facts. It is whether it interprets them better than the price already does.

That is a much higher bar. A model can correctly predict the winner and still make a bad bet if the odds were too short. It can also predict fewer winners but earn more by finding occasional mispriced outcomes.

This is the difference between forecasting and betting. One asks who is likely to win. The other asks whether the offered price is wrong.

More research did not automatically help

A separate 2026 live-trading benchmark placed AI agents into real prediction markets with capital. Returns on Kalshi ranged from -16.0% to -30.8% during the main cohort, while the same models performed much better on Polymarket. The researchers found that research volume itself had no clear relationship with returns; platform design and execution mattered. [2]

That result is awkward for the “AI will research harder than humans” story. Gathering more material can produce longer reasoning without improving the decision.

A model must also know when not to trade, how much to risk, whether liquidity is sufficient and whether its claimed probability is calibrated.

What AI did reveal

The World Cup benchmarks are useful precisely because they did not produce a clean superhero ending.

They show that:

• Fluent reasoning can accompany ordinary forecasts
• Similar top picks can hide different calibration
• A correct prediction can still be a poor trade
• Market prices remain formidable baselines
• Staking and execution affect returns as much as the headline pick
• Self-criticism after a loss varies dramatically between models

AI may become valuable as an analyst, a record keeper or a challenger to human assumptions. That is different from an automatic profit machine.

The public version will still look better

Expect social feeds to show the model that picked the champion, the exact upset or the winning accumulator. The full ledger will receive less attention.

This is the same selection problem seen with human tipsters. A spectacular prediction creates a post. A calibrated 51% estimate that loses quietly does not.

The proper test is prospective: publish every probability and every stake before kickoff, then score the complete set.

My view

The 2026 World Cup did not prove that AI cannot beat sportsbooks. One tournament and a few models cannot settle that question forever.

It did prove that intelligence-looking text is cheap compared with a repeatable edge. The market already contains models, experts, money and incentives. Arriving with a web browser and a confident paragraph is not enough.

AI can make betting analysis faster. It can also make overconfidence faster. Until a system beats closing prices and survives transparent out-of-sample testing, treat it as a research assistant, not a licence to increase the stake.