Can AI Beat the Sports activities Betting Market? 8 of the Prime Fashions Tried

General Reasoning simply gave frontier AI its worst report card but. Eight prime fashions, together with Claude, Grok, Gemini, and GPT-5.4, have been every given a digital bankroll and requested to construct a machine studying betting technique throughout a full 2023-24 English Premier League season.

Each single one misplaced cash. A number of went utterly bankrupt.

The benchmark is named KellyBenchnamed after the Kelly criterion, a 1956 formulation that tells you precisely how a lot to guess when you’ve gotten an edge over the market. Each mannequin might recite the Kelly formulation. None of them might really use it.

xAI’s Grok 4.20 failed all three runs, going absolutely bankrupt in a single, forfeiting mid-season within the different two. Google’s Gemini Flash forfeited two of three runs after inserting a single wager of roughly £273,000 on a three-percentage-point historic win-rate edge—and shedding it. Claude Opus 4.6, Anthropic’s finest mannequin, misplaced 11% on common and someway got here out trying just like the accountable grownup within the room.

In actual fact, the analysis paper mentions that the previous Dixon-Coles from the late Nineteen Nineties outperformed a lot of the frontier fashions evaluated — ending forward of six out of eight, even with restricted information.

“Dixon-Coles is an outdated 2000s baseline which doesn’t utilise all obtainable information or account for non-stationarity in a principled approach,” the researchers observe. “It’s due to this fact much more shocking that many frontier fashions, akin to Gemini 3.1 Professional, are unable to beat or match it on KellyBench.

This issues past soccer. Earlier this 12 months, AI benchmarks confirmed that Claude could dominate business simulations by way of price-fixing, cartel agreements, and strategic deception.

That call-making course of concerned static competitors, restricted opponents, clear scoring, and so forth. KellyBench is the other: 120 matchdays, always shifting information, a market that will get smarter each week, and promoted groups with zero historic information.

The researchers name the core downside a “knowledge-action hole.” It’s precisely what it appears like.

Enterprise selections are principally based mostly on mounted circumstances whereas sports activities betting is a extra fluid and mutable market, which makes issues tough for these fashions. “KellyBench requires brokers to take care of coherent intent throughout probably hundreds of sequential selections, monitor the implications of these selections, and shut the loop between statement and motion,” researchers argue.

We’re not there but, clearly.

The fashions might articulate the correct technique, diagnose when one thing was damaged, and determine the reason for their losses, however then didn’t confirm their code really carried out what they deliberate, failed to note when execution diverged from intent, and didn’t act on their very own findings.

GLM-5 wrote three separate self-critique paperwork throughout its run. Every one accurately recognized that its hardcoded 25% draw fee and overestimation of dwelling benefit have been destroying its returns. At one level, with its bankroll round £44,200, it famous that its predicted 40% dwelling win fee was solely hitting 30% in actuality. It by no means modified the code. It stored betting the identical approach till the cash was gone.

Kimi K2.5 did one thing arguably extra spectacular and extra tragic. It wrote a mathematically appropriate fractional Kelly staking operate—the correct formulation, correctly structured. Then it by no means known as it. A formatting bug induced the mannequin to ship a damaged bash command roughly 50 occasions in a row. Its reasoning famous the issue. It then despatched the an identical damaged command once more. An unintended £114,000 guess—98% of its remaining bankroll—on a Burnley versus Luton match completed the job.

GPT-5.4 was probably the most methodical. It spent 160 device calls constructing fashions earlier than inserting a single guess, then calculated that its log-loss (0.974) was barely worse than the market’s (0.971) and concluded it had no edge. It spent the remainder of the season inserting penny bets to protect capital. Sound reasoning.

OpenAI’s mannequin misplaced 13.6% on common. One seed alone price roughly $2,012 to run.

Ross Taylor, Common Reasoning’s CEO and former Meta AI researcher, told the Financial Times that almost all AI benchmarks function in “very static environments” that bear little resemblance to the true world. “There’s a whole lot of pleasure about AI automation, however there have not been many makes an attempt to guage AI in long-term, real-world environments,” he stated.

The Common Reasoning group didn’t instantly reply to a request for feedback by Decrypt.

To measure technique high quality past uncooked returns, the researchers constructed a 44-point sophistication rubric with quantitative betting fund consultants—protecting characteristic growth, stake sizing, non-stationarity dealing with, and execution. Claude Opus 4.6 scored highest at 32.6%. Lower than a 3rd of accessible factors. On the most effective mannequin.

Increased sophistication scores considerably predicted decrease chapter charges (p = 0.008) and correlated with higher general returns. The fashions are usually not failing as a result of the market is unbeatable. They’re failing as a result of they aren’t utilizing what they’ve.

This suits a sample. Research published last year discovered AI fashions develop one thing resembling playing habit when informed to maximise rewards—going bankrupt as much as 48% of the time in simulated slot machine exams. A separate real-money crypto trading competition discovered the identical reliability issues over prolonged intervals.

One of the best-performing mannequin averaged a ultimate bankroll of £89,035—a web lack of £10,965 on a normalized £100,000 beginning stake. Gradient boosting, fractional Kelly staking, months of Premier League soccer, state-of-the-art efficiency… all simply to get rekt.