Safer Gambling WeekWe tested the popular betting systems on real races. See what they madeSet a deposit limit before your next bet. Here is howNo licence means no protection. Check any bookmaker in a minuteA time-out takes a minute to set and can last a day or six weeksRaising the stake after a loser does not win it back. We ran the numbersBig Saturday of racing: decide what today can cost before the first raceWorried about your betting, or someone else's? Free help is a call away

StableBet
The AI Lab · Silicon Tipster League · Google

Gemini: horse racing tips, tested

Google's Gemini is the model most people have already met through Search, the voice behind the AI answers that sit above the results. Here it steps out of the search box and onto the racecard, picking one horse a race against four rival AIs, in the open.

#2 of 7 in Who wins?#5 of 7 in Which price is wrong?Today's picks ↓The League scoreboard ↗Research, not tips
Professor Furlong and Pascal in the StableBet AI Lab

Gemini's scorecard

8,897 picks settled

AI Picks the Winner

−4,897 pts(-10.6% return)

Never sees the odds, reads the race itself.

Points risked
46,140
Points back
41,243
Bets
4,614
Win%
20%

AI Finds the Value Bets

−9,027 pts(-21.1% return)

Sees the market's implied chance, hunts the value it missed.

Points risked
42,830
Points back
33,803
Bets
4,283
Win%
9%

Points are what a £10 bet on each pick would have won or lost: a 33/1 winner scores 330, a loser −10. No money is involved.

Went its own way: 132 races where Gemini was the lone dissenter while the rest of the field agreed on a different horse.

Correction, 24 September 2026: The going was missing from our morning cards, 3 to 23 September. Read the note

Gemini's running points total: Who wins? v Which price is wrong?

One line per job, a notional £10 a pick, settled to Starting Price and shown in points. The gap between them is the interesting bit: is Gemini better at naming winners or at spotting a wrong price?

−9,170−6,582−3,994−1,4071,1810 points, break-even07-0507-2908-2109-1310-0610-29Odds hidden−4,897 ptsOdds shown−9,027 pts
Odds hidden−4,897 ptsOdds shown−9,027 pts
Running points total on a notional £10 a pick, settled to industry SP. Research, not tips.

How Gemini's last calls went

Settled picks, newest first, with the one-line reason Gemini logged before the off and what the race did with it. Wins and losses alike, scored at Starting Price.

RaceJobPickConf.ReasonResult
Newcastle19:40Which price is wrong?South ShoreMediumCowell sprinter, good draw, recent form suggests better than market.Lost · SP 15 · −10 pts
Newcastle19:40Who wins?Curious RoverMediumHollie Doyle aboard, consistent form, good draw for a strong run.Lost · SP 13 · −10 pts
Newcastle19:10Which price is wrong?Irish IncentiveMediumDropping in class, good jockey, could surprise at long odds.Lost · SP 19 · −10 pts
Newcastle19:10Who wins?Forever TwentyMediumHollie Doyle aboard, consistent form, good draw. Looks promising for a win.Lost · SP 5 · −10 pts
Newcastle18:40Which price is wrong?Percy's StarMediumConsistent form, good jockey, market underestimates recent improvements on all-weather.Lost · SP 29 · −10 pts
Newcastle18:40Who wins?We Never StopMediumConsistent form, good jockey/trainer combination, handles the going well.Lost · SP 4.3333 · −10 pts
Newcastle18:10Which price is wrong?Pol RogerMediumDropping in class, good draw, trainer in form, market underestimates.Lost · SP 81 · −10 pts
Newcastle18:10Who wins?LuminareHighHarry Davies takes the ride for Haggas; strong combination, good form.Lost · SP 10 · −10 pts
Newcastle17:40Which price is wrong?InkjetMediumHaggas runner, debutant with good breeding, market underestimates potential.Lost · SP 5.5 · −10 pts
Newcastle17:40Who wins?InkjetHighHaggas debutants often impress, good jockey booking, strong stable form.Lost · SP 5.5 · −10 pts
Newcastle17:10Which price is wrong?Bronze CascadeMediumDebutante from good yard, market underestimating potential improvement.Lost · SP 13 · −10 pts
Newcastle17:10Who wins?TabkirHighGosden/Havlin combination is strong, promising debut, likely to improve significantly.Lost · SP 4.5 · −10 pts
Newcastle16:40Which price is wrong?Angelic MelodyMediumConsistent form, good trainer/jockey combo, better than market suggests.Won · SP 5.5 · +45 pts
Newcastle16:40Who wins?KumharasMediumFahey runner with a good jockey, consistent form, looks ready to win.Lost · SP 15 · −10 pts
Newcastle16:10Which price is wrong?EzenliMediumDoyle booking for Watson, better than bare form suggests, market undervalues.Won · SP 34 · +330 pts
Newcastle16:10Who wins?El Ra'adyHighStrong trainer and jockey combination, good form on standard going.Lost · SP 2.2 · −10 pts
Lingfield15:40Which price is wrong?Shalaa AskerMediumCourse/distance winner, dropping in class, good draw, market underestimates.Lost · SP 15 · −10 pts
Lingfield15:40Who wins?Solar InvincibleMediumBilly Loughnane takes the ride on a consistent sprinter, good form at this track.Lost · SP 6 · −10 pts
Newcastle15:35Which price is wrong?Paddy The SquireMediumConsistent form, good jockey, market underestimates recent strong performances.Lost · SP 12 · −10 pts
Newcastle15:35Who wins?Water To WineHighGosden runner with top jockey, consistent form, looks well-placed here.Lost · SP 2.2 · −10 pts

Today's picks

Nothing on the card right now. Gemini's picks are logged each morning once the day's racecard publishes, appear here before the first off, and move up into the settled record as the results land. On a day with no British or Irish racing there is nothing to log.

How Gemini actually picks

Hand Gemini a racecard (runners, going, class, distance) and you tend to get a brisk, decisive read. This looks like a model that lands on the plausible favourite rather than digging for a contrarian shape, so we will be watching how often its one-line reason leans on the front of the market versus something it has spotted in the conditions. The interesting question is whether that search-engine reflex (surface the most-referenced, most-obvious answer) holds up over the flat run of ordinary midweek cards, or whether the races that really test it are the tricky handicaps where the obvious answer and the winner part company. Every pick and its short reason string is published, so you can read exactly how Gemini talked itself into each horse.

Here is what the numbers say, updated live as the sample grows. Reading a race blind, with the prices hidden, Gemini backs the outright favourite 22% of the time, and its picks most resemble the “back short-priced favourites” approach (backs the favourite only when it is odds-on). Its stated reasons lean on form, jockey and trainer.

How Gemini explains itself (AI Picks the Winner)Form82%Jockey44%Trainer33%Going21%Draw8%Distance7%Class7%Course2%Weight1%Value0%0%50%100%
Share of this AI's short written reasons that mention each theme, longest first (a reason can touch several, so the shares do not add to 100). The three it leans on most are picked out in gold. This describes how it TALKS about its picks; it is not why it wins or loses. Based on 4,614 reasons.
What is Gemini actually doing? (picking blind)How much its picks resemble each tested strategy, the top 8 shown.chanceidenticalback short-priced favourites0.28favourite over jumps0.15favourite on soft ground0.12back the favourite0.11each-way the favourite0.11favourite on the Flat0.09favourite in a big field0.09favourite in a small field0.09
kappa is a chance-corrected resemblance: 1 means identical selections, 0 means no more alike than chance. Gemini's picks look most like back short-priced favourites. A bar whose interval touches 0 (marked n.s.) is not distinguishable from chance.

Show it the market and it changes character: its price picks swing toward the “back the outsider” approach, hunting the overlooked rather than the well-fancied, so they win less often and pay more when they do. Research, not tips.

What is Gemini actually doing? (shown the market)How much its picks resemble each tested strategy, the top 8 shown.chanceidenticalback the outsider0.16each-way a longshot0.14our own value strategy0.04top-rated in handicaps−0.01n.s.back the top-rated horse−0.03favourite in a big field−0.05back the second favourite−0.06our model's top pick−0.07
kappa is a chance-corrected resemblance: 1 means identical selections, 0 means no more alike than chance. Gemini's picks look most like back the outsider. A bar whose interval touches 0 (marked n.s.) is not distinguishable from chance.
What Gemini weighs (AI Finds the Value Bets)
← leans weaker0 = random pickleans stronger →−0.5σ−0.4σ−0.3σ−0.2σ−0.1σ0+0.1σ+0.2σ+0.3σ+0.4σ+0.5σstandard deviations within the racemodel-vs-market edgen=4,283course recordn.s. · n=1,420official ratingn=3,246our model's viewn=4,283trip recordn=2,810recent formn=3,928last-time-out finishn=3,921the market's viewn=4,283
Each bar is how far this AI's picks sit from a random runner on that measure, in standard deviations, within the race; a bar whose interval crosses zero is not distinguishable from chance. Some measures only cover part of the card (n), and bars marked n.s. lean no clearer than a coin toss.

See how all five AIs rank → · Read the full write-up on how Gemini picks →

About Gemini

Gemini is Google's family of AI models, and it is the one most people have already used without quite noticing, wired into Google Search and powering the AI Overviews that increasingly sit above the blue links. Its whole design brief is breadth: give a broad, usable answer to almost any question in the time it takes a page to load. That reach across an enormous sweep of the web is its signature strength.

That heritage tends to shape its character on the board. Gemini's instinct is recall and pattern-matching at scale, reaching for the answer that fits what it has seen, quickly, rather than agonising over a contrarian read. Watching it size up a racecard is a fair test of how far broad, confident, wide-angle reasoning gets you when the going is soft and the field is competitive, which is a different question from the one the other four models answer.

None of that was built for horse racing, and that is exactly the point of putting it here. Gemini has no private form book and no edge on the market: it gets the same racecard as everyone else and has to make one call. What we get to watch is whether Google's search-shaped instinct reads a race the way it reads a query: straight to the obvious, or with an angle of its own.

Who wins? v Which price is wrong?

Gemini's search heritage leans towards surfacing the consensus answer, which makes its two tests well worth comparing. In AI Picks the Winner it never sees the odds and has to read the race off the conditions alone, a real test of whether its broad recall becomes a view of its own. In AI Finds the Value Bets it also sees the market's implied chance for each runner, and here is the honest worry: a model shaped to find and echo the most-referenced answer may simply defer to the favourite once it can see one, sharpening nothing and just copying the board. If informed Gemini and the market move as one, that tells us something; if its winner-picking run quietly reads races differently, that is the more interesting Gemini to watch.

What to watch on Gemini's board

  • Whether Gemini's search-shaped instinct sends it straight to the market favourite or lets it back a live outsider
  • How its blind and informed arms diverge: does seeing the price make it sharper, or just make it copy the board?
  • The races where Gemini 'went its own way', the lone dissenter while ChatGPT, Claude, Grok and DeepSeek all agreed on another horse
  • The reason strings: whether its one-liners cite the conditions or just restate the obvious

The rest of the field

Gemini is one of 9 on the board. See how the others are reading the same races:

Questions about Gemini tips

Is Gemini good at horse racing tips?

Gemini is very good at giving a fast, plausible answer, which is not the same as being good at making money on horses. In the Silicon Tipster League it reads the same racecard as everyone else with no form-book edge and no view of the Starting Price. Treat its picks as a public experiment in how Google's model reasons, not as betting advice.

Can I make money following Google's Gemini racing picks?

No, and you should not treat them that way. Naming a likely winner is not the same as being paid enough when it lands. The betting market is a ruthlessly efficient forecaster, and across tens of thousands of races no method the Lab has tested beats the bookmaker's margin. Gemini's picks are here as research and entertainment, for interest, not profit.

Which model powers Gemini in the Silicon Tipster League?

It runs on a current Google Gemini model, given nothing but the racecard, the same runners, going, class and distance the other four AIs see. Every pick is logged before the off, settled at industry Starting Price and scored in points as a notional £10 bet, so what you read on the board is Gemini's real, unedited call.

Should I bet based on what Gemini picks?

The Lab's honest answer is no. This is a data experiment, not a tipping service. If you do choose to bet, that is your decision alone: betting is for over-18s, for entertainment, and never a way to make money. Please gamble responsibly, and treat Gemini's calls as something interesting to watch, not to follow.

This page sits inside the AI Lab, where we test the popular betting systems against the market and ask the bigger question in does following an AI tipster work?

Gamble responsibly. This page is research and entertainment, not betting advice, 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+.