Every frontier AI model tested against a real Premier League betting season lost money.
Not most of them. All eight. Several went bankrupt.
The one that held up best still finished down £10,965 on a £100,000 starting stake.
And the model that outperformed six of the eight was built in the 1990s.
The hype around AI sports betting has outrun the results by a wide margin.
What KellyBench – a new benchmark from General Reasoning – actually measured was not whether AI can understand betting strategy. It can.
The critical question is whether it can execute that strategy consistently across hundreds of sequential decisions under real market pressure.
The answer, right now, is no. This piece breaks down exactly what happened, why it happened, and what it means for bettors evaluating AI tools in 2025.
What KellyBench Actually Tested – and What It Found
General Reasoning, an independent AI safety and benchmarking firm led by CEO and former Meta AI researcher Ross Taylor, named KellyBench after the Kelly criterion – a 1956 formula that tells you exactly how much of your bankroll to stake when you have a genuine edge over the market.
The setup was straightforward: eight frontier models, including Claude Opus 4.6, GPT-5.4, Gemini Flash, and Grok 4.20, were each given a virtual £100,000 bankroll and asked to build a machine-learning betting strategy across all 120 matchdays of the 2023-24 English Premier League season.
The critical clarification: this was not prompt-based AI picking games. What it actually tested was whether AI agents could write functional code, build working prediction models, size bets correctly using Kelly staking, monitor their own performance, and adapt when the strategy was failing – all autonomously, over months of simulated time.
Every single model failed to turn a profit. The best performer averaged a final bankroll of £89,035. Dixon-Coles, a Poisson-based goal model from the late 1990s that does not account for non-stationarity and does not use all available data, outperformed six of the eight frontier models. Ross Taylor told the Financial Times that most AI benchmarks operate in “very static environments” that bear little resemblance to the real world. KellyBench was designed to be the opposite of that.
Why Sophisticated AI and Sports Betting Are a Complicated Match
The researchers identified the core failure as a “knowledge-action gap.” These models could articulate correct strategy with precision. They diagnosed their own problems in real time. They just couldn’t close the loop between knowing and doing.
GLM-5 is the cleanest case study. It wrote three separate self-critique documents during its run. Each one correctly identified that its hardcoded 25% draw rate and overestimation of home advantage were destroying returns – at one point noting its predicted 40% home win rate was only hitting 30% in reality, with its bankroll sitting at £44,200. It never changed the code. It kept betting the same broken model until the money was gone.
Kimi K2.5 did something arguably worse. It wrote a mathematically correct fractional Kelly staking function – the right formula, properly structured. Then a formatting bug caused it to send a broken bash command roughly 50 times in a row. Its reasoning log noted the problem. It sent the identical broken command again. An accidental £114,000 bet – 98% of its remaining bankroll – on Burnley versus Luton finished the run.
Grok 4.20 failed all three runs: bankrupt in one, forfeited mid-season in the other two. Gemini Flash forfeited two of three runs after placing a single wager of roughly £273,000 on a three-percentage-point historical win-rate edge and losing it. These are not close calls. This is structural execution failure at scale.
The math hardened fast. As a 2024 preprint by Chirkunov and Füzesi on using ChatGPT for football betting found, naïve prompt-based picks produced negative ROI once realistic odds and bookmaker margins were applied – even when the model could explain value betting concepts fluently. KellyBench confirmed the same pattern at a far more rigorous level of testing.
Bettors treating AI as a shortcut to edge are making a category error. The models are not failing because the market is unbeatable. They are failing because they cannot reliably run what they built.
Where AI Creates Genuine Value in Betting – and Where It Doesn’t
The edge is not in autonomous AI agents managing your bankroll. That much is settled.
What the KellyBench sophistication rubric did show is instructive. Researchers built a 44-point scale covering feature development, stake sizing, non-stationarity handling, and execution quality. Claude Opus 4.6 scored highest at 32.6% – less than a third of available points, but still the best of the group. Higher sophistication scores significantly predicted lower bankruptcy rates (p = 0.008) and correlated with better returns. The direction of the relationship is real even if the absolute performance is not.
That points to where AI tools genuinely help today: as decision-support systems, not autonomous agents. A human who understands the Kelly criterion and uses AI to build, test, and interrogate a prediction model – rather than handing the whole process to the agent – gets the analytical upside without the execution failure risk. The AI handles data processing and model construction. The human handles verification and stake decisions.
The specific use case that works: using AI to surface non-obvious statistical relationships in large datasets (player tracking data, weather, rest days, referee tendencies) and then applying disciplined manual staking. This is exactly how sophisticated sports bettors already use quantitative tools, and it is a long way from “AI beats the book.”
Specialized quantitative models with human oversight have shown modest historical edges – typically under 5-7% ROI in backtests – though these generally assume frictionless execution. Once major bookmaker margins and account restrictions are factored in, sustainable edge shrinks further. The market is not soft. AI betting strategy applied to World Cup 2026 accumulators shows how even carefully constructed AI-assisted models must account for those structural headwinds before a single bet is placed.
Best Bet Structures for AI-Assisted Handicapping
Given what KellyBench proved and what the broader research landscape supports, here are the specific structures that represent actual positive-EV use of AI tools – not fantasy, not hype.
Structure One: AI-Generated Feature Sets, Human Staking Decisions
Use AI to build and stress-test a prediction model. Validate its outputs against closing lines from sharp books (Pinnacle, Circa). Only bet when your model’s implied probability exceeds the closing line by at least 3 percentage points – that gap is your edge signal. Size flat at 1-2% of bankroll per bet, not Kelly, until you have 500+ bets of track record to estimate true edge variance.
\p>The Bet: This structure is not a specific pick – it is the framework. Applied to high-volume, high-data environments like March Madness first-round underdogs, where model inputs are most reliable and the market is least efficient, it has the highest expected hit rate.
Structure Two: Dixon-Coles Baseline Calibration
The KellyBench result was not an endorsement of Dixon-Coles as a viable current tool – it was an indictment of how badly frontier AI executed by comparison. A disciplined bettor using any calibrated Poisson-based goal model, verifying outputs manually, and applying fractional Kelly (half-Kelly at most) is better positioned than any of the tested AI agents. The Bet: fractional Kelly staking (0.25x–0.5x) on markets where your model shows 4%+ edge over closing line, maximum two legs in any parlay.
Structure Three: Line-Shopping Automation
The one place AI tools deliver unambiguous value today is real-time line comparison across books. This does not require frontier model reasoning – it requires reliable execution of a simple task. Automated line-shopping across Pinnacle, DraftKings, FanDuel, and Caesars for the same market is worth 0.5-1.5% per bet in recovered juice. That is real, compounding, and available right now.
Risks and What the Models Cannot Predict
The KellyBench results specifically exposed failure modes that matter for any bettor evaluating AI tools. Non-stationarity is the primary one: promoted teams with zero Premier League historical records broke every model that relied on historical win-rate baselines. GLM-5’s 40% predicted home win rate against a 30% actual is a direct consequence of applying a static model to a dynamic market.
Market adaptation is the second structural risk. Major sportsbooks use machine-learning models for live pricing and risk management, and the UK Gambling Commission has noted that operators increasingly use automated risk tools to detect and limit sharp bettors. Even if an AI model found a genuine edge, monetizing it through standard retail books would face account restriction before the sample size became meaningful.
The gambling-addiction behavior pattern is real and documented. Research published last year found AI models develop something resembling addiction-like behavior when optimizing for reward – going bankrupt up to 48% of the time in simulated slot machine tests. Grok 4.20’s complete bankruptcy and Gemini Flash’s £273,000 single-wager decision fit that pattern exactly. Any autonomous AI agent given a bankroll and a reward signal is operating near that failure mode by design.
Practical rule: never delegate stake sizing to an AI agent. Verify every function that touches bet size before it runs. If you cannot read the code, you cannot trust the output.
Bottom Line
AI cannot beat the sportsbook autonomously – not in 2025, not with any of the eight frontier models tested, and not across a full season of real market conditions. Every model lost money. A 1990s statistical baseline outperformed six of them. The failure was not intelligence; it was execution reliability across hundreds of sequential decisions in a fluid, adversarial market.
The specific recommended approach: use AI as a feature-generation and model-building tool, apply human verification at every step, stake fractional Kelly at 0.25x–0.5x, and only act when your model shows 3-4% edge over closing line. High-volume data environments with public market inefficiencies – first-round tournament games, early-season league markets – are where that edge is most likely to exist.
Do not hand a bankroll to an AI agent. Build the model, verify the code, and make the call yourself.
