Dan King's Galway Double: What Our AI Model and Five Chatbots Actually Predicted

The Plate: nobody was anywhere near it

Twenty-five runners were priced for the Tote Galway Plate. Our model made Buddy One its most likely winner and gave him a shade under 6 in 100. That is the important number, and it is not a selection. In a race like this the model's honest position is that no horse is likely to win, and its estimates ran in a tight band from about 6 in 100 down to 3 in 100 across the whole field.

King Alexander, the horse that actually won, sat eighteenth of the twenty-five on the model's list, at roughly 3.6 in 100.

The chatbots did no better, and they did something more revealing. Of the thirteen selections logged for the race, six landed on the same horse, Funiculi Funicula, the market favourite. Four of those six were blind picks, made without seeing a price at all. Only six different horses were chosen out of a field of twenty-five.

Entrant	Blind pick	Informed pick
ChatGPT	Buddy One	Funiculi Funicula
Claude	Funiculi Funicula	Jesse Evans
Gemini	Nowwhatdoyouthink	King Of Kingsfield
Grok	Funiculi Funicula	Perceval Legallois
DeepSeek	Funiculi Funicula	Funiculi Funicula
Our model	not entered blind	Buddy One

Not one of them named King Alexander. Funiculi Funicula did not win either.

The clustering is the story. Asked to find a winner in a twenty-five-runner handicap, several different models independently reached for the same obvious horse.

The Hurdle: one hit, and it was not a confident one

Twenty-one runners were priced for the Guinness Galway Hurdle. Gemini, shown the market's odds, picked Putapoundinthejar. It won at 10/1. That is the only winner among the twenty-six selections logged across the two features, and it deserves to be reported plainly rather than dressed up.

Gemini put the horse's chance at 6 in 100 and logged its confidence as medium. Twelve other selections went elsewhere. Three of the five blind arms independently chose the same horse, Murcia, which did not win. Read as a whole, the round looks like thirteen guesses at a lottery, one of which came in.

Our own model had Putapoundinthejar eighth of twenty-one, at about 5 in 100. It was not blind to him, but it was not pointing at him either.

The model's own top two are worth a mention, because they were closer than the raw result suggests. It rated Ragmans Corner the most likely winner and Gameball second. Ragmans Corner finished fourth. Gameball finished third. Both were double-figure prices, and both ran into the frame in a race the market had shaped very differently, with El Cairos sent off favourite.

That is the pattern worth noticing across both days. The model was not finding winners. It was, in the Hurdle at least, pointing at the right sort of horse at the right sort of price, and then being beaten by one it rated a little lower. In a twenty-one-runner handicap that is roughly what a well-calibrated set of numbers should look like.

What two races like these actually cost

The reason nobody found King Alexander is not that the model is broken or that the chatbots are stupid. It is that these races are close to unreadable, and they are priced accordingly.

Add up the starting prices of every runner in the Galway Plate and the book comes to about 164 per cent. In the Galway Hurdle it was about 146 per cent. A fair book adds to 100. Everything above that is the margin built into the prices, and on those two races it works out at roughly 39 pence in every pound staked on the Plate and 31 pence on the Hurdle.

That is the number that decides most punters' week, and it is far larger than any edge a tipster, a model or a chatbot is going to find in a twenty-five-runner handicap. Our model's most confident opinion in either race was under 7 in 100. When the best available estimate is that nothing is likely, the margin is doing all the work.

None of which makes the races less enjoyable, and Dan King's week is a genuinely rare piece of history. It just means the honest way to treat a Plate or a Hurdle is as a spectacle you might have a small interest in, not a puzzle waiting to be solved.

Saturday's Stewards' Cup at Goodwood is built the same way: a huge field, a wide-open market and a book to match. Our fair prices and the margin on every race publish on the racecards each morning, and the AI Lab shows how the model has actually scored over time, losers included. Read them as a measure of how much the race is costing you, not as a selection.

Research, not tips. If you bet, stake only what you can afford to lose. BeGambleAware.org. 18+.