AI Picks the Winner
Can AI pick winners on form alone?
8,422 picks settled · £1 level stakes to SPThe live leaderboard
All settled picks, ranked by return on investment.
| # | AI | Staked | Returned | Profit | Bets | Win% | Backed the fav | Profit trend |
|---|---|---|---|---|---|---|---|---|
| 1 | ClaudeAnthropic | £1,397 | £1,301 | −£95.81(-6.9%) | 1,397 | 22% | 27% | |
| 2 | The FavouriteThe market | £1,364 | £1,256 | −£108(-8.0%) | 1,364 | 30% | 100% | |
| 3 | GrokxAI | £1,445 | £1,245 | −£200(-13.8%) | 1,445 | 23% | 28% | |
| 4 | GeminiGoogle | £1,445 | £1,237 | −£208(-14.4%) | 1,445 | 21% | 23% | |
| 5 | DeepSeekDeepSeek | £1,326 | £1,081 | −£245(-18.5%) | 1,326 | 22% | 29% | |
| 6 | ChatGPTOpenAI | £1,445 | £1,172 | −£273(-18.9%) | 1,445 | 22% | 30% |
Millions of people now ask an AI chatbot for a horse racing tip. AI Picks the Winner is that moment, run properly: each AI is handed the racecard (the runners, the going, the class, the trip) with one thing withheld: the odds. No prices, no market, no favourite pointed out. It has to read the race on the form alone, which makes this the purest test of the question everyone actually means when they ask “can AI pick horses?”
One competitor of our own runs here: The Favourite, which just backs the morning market leader in every race. No analysis, no reading, the bet a punter with nothing else falls back on, and the baseline every tipster in this test has to beat. (Our own engine enters AI Finds the Value Bets twice, as Stablebet Edge and the Stablebet Model: it is built on market data, so calling it blind would be dishonest.)
Research, not tips. Settled honestly to industry SP, wins and losses alike. 18+ · please gamble responsibly.
What the numbers say
The read on 44 days and 16,055 settled picks, worked out live from the board.
Every AI finds the winner in about 21.3%–22.6% of races, within a whisker of each other, and every one still loses. The market prices those winners too short. Naming the likely winner is not the same as being paid enough for it.
They back the market favourite 23%–30% of the time, and went against a unanimous field only 16–57 times each. Most of the time, asking an AI is close to just backing the favourite.
Not enough: none beats the bookmaker's margin, and nobody is in profit.
Return per £1 staked, settled to Starting Price. The three house entrants are shown alongside the chatbots: The Favourite (in AI Picks the Winner), plus Stablebet Edge and the Stablebet Model (both in AI Finds the Value Bets). Research, not tips.
The difference: AI Picks the Winner vs AI Finds the Value Bets
Each chatbot runs twice on the same races: once blind (odds hidden, picking the winner on its own read) and once informed(the market's implied chances on show, hunting the value the market has missed). The gap between the two records is what the market's view (and the switch from winner-picking to value-hunting) does to each model's returns. On average, the informed run moved returns by £4.68 per £100 staked.
| AI | Blind ROI | Informed ROI | Winner-vs-value gap | Backed the fav (blind → informed) |
|---|---|---|---|---|
| Claude | -6.9% | -23.2% | −£16.30 | 27%→0% |
| Grok | -13.8% | -23.7% | −£9.90 | 28%→2% |
| ChatGPT | -18.9% | -24.6% | −£5.70 | 30%→0% |
| DeepSeek | -18.5% | -14.4% | £4.10 | 29%→23% |
| Gemini | -14.4% | -10.0% | £4.40 | 23%→0% |
Blind, the AIs back the market favourite some of the time. Show them the odds and they pile onto it: the informed column above sits far higher. The “informed” edge turns out to be little more than backing the jolly, and the favourite loses too.
Finding the winner and getting paid enough for it are two different things. The market prices the likely winners too short, so picking more of them does not turn into profit.
The gap = a competitor's blind return minus its informed return, at £1 level stakes to Starting Price. The two tests ask different questions (blind picks the winner; informed hunts value; its record restarted 17 July 2026 when we fixed the question), so read this as the cost of the whole informed way of playing, on early samples. Treat any single positive arm as noise, not a system. Research, not tips.
Two tests, and why
Every AI in the league runs twice on the same races, and we ask each version the question that fits what it can see. Here in AI Picks the Winner, the odds are hidden, so the fair question is the direct one: who wins? In AI Finds the Value Bets, the market's implied chance for every runner is on the card (asking “who wins?” there would answer itself), so the question becomes: where is the value the market has missed? Two experiments, kept deliberately apart, and the gap between their records is what the market's view does to a machine's judgement.
How we test it
Every race in Britain and Ireland, each AI is handed the racecard (the runners, the going, the class, the distance) and asked for one thing: a pick, and a one-line reason. That pick is timestamped and written to the record before the race is run, so there is no way to sneak a look at the result.
Once the race is settled we grade the pick at industry Starting Price, at £1 level stakes, with fallers and non-finishers counted as the losers they are, exactly the convention we use for every other study in the Lab. Strike rate is how often the pick wins; return to SP is what £1 a time would have done.
Three competitors of our own run alongside the chatbots: one here in AI Picks the Winner, two in AI Finds the Value Bets. The Favouriteruns here: it mechanically backs the morning market leader, no analysis at all, and every row on this board answers to it: a tipster that can't beat just-back-the-favourite hasn't added anything. Our own engine enters AI Finds the Value Bets twice, under two rules. Stablebet Edgebacks the runner it rates furthest above the market's implied chance, its biggest edge on the card; the Stablebet Model backs the horse it rates most likely to win. Same numbers, two answers, and the gap between their records is its own experiment. The engine is trained on market data, so it sees the market by construction; calling either entry blind would be dishonest.
A note on honesty: we test a current, capable model from each lab (GPT-4.1, Claude Sonnet, Gemini, Grok and DeepSeek) called through each provider's API, and every pick is published with the exact model id that made it, so you can check us. Not the priciest reasoning flagships that burn hidden “thinking” tokens, just models chosen so every race, every day, is affordable to log and fully reproducible. And one day of racing proves nothing; the value is in the sample building over weeks, which is why we started it in public on day one.
The competitors
The one most people actually ask for a tip. Runs on OpenAI's GPT models.
See ChatGPT's picks & record →Google's flagship, and the model behind many of the AI answers you see in search.
See Gemini's picks & record →Anthropic's model, known for careful, measured reasoning.
See Claude's picks & record →xAI's model, wired into the firehose of real-time chatter.
See Grok's picks & record →The efficient challenger from China that shook up the field on cost.
See DeepSeek's picks & record →Our own racing engine, backing the runner whose price sits furthest above the model's view of its true chance: the biggest edge on the card, however unfancied. Built on market data, so it competes in AI Finds the Value Bets.
See Stablebet Edge's picks & record →The same engine, backing the horse it rates most likely to win each race. One model, two rules: this entry backs its winner, while Stablebet Edge chases the value. The gap between their records is the experiment.
See Stablebet Model's picks & record →No opinions, no analysis, just backs the morning favourite in every race. The baseline for AI Picks the Winner, and the bar every tipster has to beat.
See The Favourite's picks & record →What it will probably show
We are curious which AI comes out on top, but we would be surprised if any of them turns a profit over a real sample, and we will say so plainly if the data proves us wrong. The reason is not that the models are stupid; it is that the betting market is one of the most efficient forecasters ever built. By the time a price is set, thousands of people have already bet everything they know into it.
So the likely story is the Lab's whole thesis in miniature: an AI can be really good at naming the most likely winner and still lose you money, because naming the winner is not the same as being paid enough when it happens. That is a useful thing to prove in public, because "just ask ChatGPT for a tip" is advice a lot of people are quietly following.
Questions
Which AI is the best horse racing tipster?
That is exactly what this experiment measures, and honestly, nobody knows yet, because it starts from scratch. Each day we log every model's pick before the race and settle it at Starting Price. The AI Picks the Winner board above is the live answer as it builds. Come back and watch it move.
What does AI Picks the Winner measure?
Whether an AI can read a horse race from the form alone, with no odds to lean on. Each model sees the racecard but not the prices, picks one winner before the off, and we settle it at Starting Price. It isolates genuine race-reading from simply echoing the market.
Why hide the odds from the AI?
Because the market price already contains almost everything knowable about a race. Show an AI the odds and it can score well just by leaning on the favourite, which tells you nothing about its own judgement. Hiding the odds forces it to form a view of its own, which is the interesting thing to measure. Its sibling test, AI Finds the Value Bets, runs the opposite experiment: market shown, value demanded.
Should I follow ChatGPT's betting tips?
This page is built to answer that with real numbers rather than opinion. Our wider research is blunt: no selection method we have tested beats the bookmaker's margin over a real sample, and an AI naming a likely winner is not the same as being paid enough when it wins. Treat any AI tip as entertainment, never a way to make money.
Do you bet real money on these picks?
No. Every pick is settled on paper at Starting Price, the fairest and hardest-to-flatter convention. This is research, not a tipping service, and nothing here is a signal to stake.
Gamble responsibly.This page is research and entertainment, not betting advice. No AI here beats the bookmaker's margin, and nothing on it is a signal to stake. Betting should never be a way to make money. If it is affecting you or someone you know, free and confidential support is at BeGambleAware.org. 18+.