StableBet
StableBet
Professor Furlong and Pascal at the AI Lab
THE AI LAB
LAB NOTES · LEAGUE ROUNDUP

Champions Day 2026: Stablebet Edge Finds Mount Atlas at 33/1, and the AI League's Day at Ascot

Stablebet Edge, the value-hunting entry from our own racing model, picked out Mount Atlas before the British Champions Long Distance Cup at Ascot, and he won at 33/1. That one result made the Edge the top entrant in the Silicon Tipster League on Champions Day, with a return of +385.7% on its stakes for the afternoon, against -15.4% across its season.

Every entrant over the 7 races of British Champions Day 2026, one unit a pick, settled at starting price. Voids are excluded from both the bet count and the return.
#EntrantTestBetsWonStrikeP/LReturnBacked fav
1Stablebet EdgeAI Finds the Value Bets7114.3%+£27.00+385.7%0%
2ChatGPTAI Picks the Winner7228.6%+£22.00+314.3%14%
3ClaudeAI Picks the Winner7114.3%+£16.00+228.6%43%
4DeepSeekAI Picks the Winner7114.3%+£16.00+228.6%43%
5Stablebet ModelAI Finds the Value Bets7228.6%−£0.90−12.9%86%
6GrokAI Picks the Winner7114.3%−£1.00−14.3%29%
7GrokAI Finds the Value Bets7114.3%−£1.00−14.3%0%
8DeepSeekAI Finds the Value Bets7114.3%−£1.00−14.3%0%
9The FavouriteAI Picks the Winner7114.3%−£4.90−70.0%100%
10StableBet RatingAI Picks the Winner100.0%−£1.00−100.0%0%
11GeminiAI Picks the Winner700.0%−£7.00−100.0%0%
12ChatGPTAI Finds the Value Bets700.0%−£7.00−100.0%0%
13GeminiAI Finds the Value Bets700.0%−£7.00−100.0%0%
14ClaudeAI Finds the Value Bets700.0%−£7.00−100.0%0%

14 entrants, 92 bets, +£43.20 overall at +47.0%. 4 finished the meeting ahead. the entrant backs the horse heading the market on the 09:05 card, settled at SP - NOT the horse that started favourite. See systems[key=favourite] for the SP-favourite strategy; over this meeting the two differ by more than the meeting's whole margin.

The Edge works to a single rule. In every race it finds the runner whose price sits furthest above our model's estimate of its chance, and backs it, however unfancied. On the morning card the model rated Mount Atlas at a little more than twice the chance the market was giving him, and no other runner in the race showed a bigger gap. Andrew Balding's stayer, ridden by PJ McDonald, beat Queenstown by three-quarters of a length, with the 7/4 favourite Highwayman a further three-quarters back in third.

That was the Edge's only winner. Its other picks on the card, every one of them priced at 18/1 or longer, all lost. That is how the entrant is built to bet: it backs horses the market has largely written off, so it loses most races and relies on a long price landing now and again.

The day's figure has to sit beside the season's, and the season's is negative. Across 3,210 settled bets the Edge has returned -15.4%, with a strike rate of 8%. A 33/1 winner moves one afternoon a long way and a record that size hardly at all. We are pleased with the pick, and it changes nothing about the record.

A good afternoon for the League, set beside its season

Across all 14 entrants the League placed 92 bets at Ascot and returned +47.0% on them. Over the season the best record of any entrant is -7.2%. Of those entrants, 4 finished the day in front. This League rarely prints a line like that, and two long-priced winners account for nearly all of it: Mount Atlas for the Edge, and Fifth Column at 22/1 in the Balmoral Handicap, which three of the blind arms had backed.

A card of 7 races is nowhere near a sample. Take away Mount Atlas and Fifth Column and nobody finishes the day ahead. So each entrant that did finish ahead belongs beside its full record:

  • Stablebet Edge: -15.4% over 3,210 settled bets.
  • ChatGPT, form only: -12.0% over 3,601.
  • Claude, form only: -7.9% over 3,553.
  • DeepSeek, form only: -15.8% over 3,479.

There is a useful yardstick for how often luck alone produces an afternoon like this. In our simulation of the same card, a pick made at random in every race averaged −10.9%, yet finished ahead in 17.8% of trials, because the odd big-priced winner drags a short run into profit. That is a property of the simulation, not a measurement of real punters, but it shows how little a single good day can tell you.

One more pattern from the day, with the same warning attached. The blind arms, shown the form with no prices, did better than the informed arms, which see the market and are asked where they disagree with it. The informed arms of ChatGPT, Claude and Gemini went through the whole card without a winner. It is one day. The question the League keeps asking is whether seeing the prices helps a model or costs it, and the season records are where that gets answered: Claude sits on -7.9% blind against -26.9% informed, and ChatGPT on -12.0% against -24.2%. The value test sets every pair side by side.

Favourites, longshots and a crowded opener

The card split the field into entrants that sided with the market and entrants that went looking for a price, and the market's leaders had a mixed afternoon.

The favourites that let people down. Kalpana started at Evens for the Fillies & Mares Stakes and finished seventh. Claude and DeepSeek on their form-only arms were with her, as were Stablebet Model and The Favourite. Gstaad was sent off the 7/4 favourite for the Queen Elizabeth II Stakes and finished eighth, carrying the form-only picks of ChatGPT, DeepSeek and Grok as well as The Favourite. Ombudsman did win the Champion Stakes at 5/4, but no entrant had him. The morning favourite in that race was Calandagan, who ran second, and five entrants were on him.

The Favourite is a morning bet. The League's baseline backs whichever horse heads the market on the 09:05 card, not the one that starts favourite. Those were different horses in three of the races: the juvenile opener, the Champion Stakes and the Balmoral. The baseline found one winner, Almeraq at 11/10 in the Sprint, and returned −70.0% on the day. Backing the starting-price favourite in every race instead returned −37.9%. Over the season The Favourite stands at -11.3%. Which favourite you mean changes the answer, so we label them separately.

Stablebet Model kept close to the market. The same engine as the Edge, it backs the runner it rates most likely to win, which on this card was usually the morning favourite. It had two winners, Almeraq and Talk Of New York at 3/1 in the QEII, and still finished a little short of level for the day. Its season figure is -17.2%.

The blind arms found Minnie Hauk and Fifth Column. ChatGPT's form-only arm had Minnie Hauk at 5/1 in the Fillies & Mares, and so did both Grok arms and DeepSeek's informed arm. In the Balmoral, ChatGPT, Claude and DeepSeek, all on form only, landed on Fifth Column. Claude and DeepSeek both cited the John and Thady Gosden and William Buick combination; ChatGPT pointed to the jockey and a progressive profile. Buick was the day's top jockey with 3 winners.

The opener showed the models herding. Seven of the arms backed Confucius in the juvenile race, nearly all quoting Aidan O'Brien and Ryan Moore. He went off at 20/1 and finished fourth.

Gemini had no winner on either arm. Nor did the informed arms of ChatGPT and Claude. StableBet Rating, our handicap-only rating, had a single bet, Persica in the Balmoral, the same horse as the Edge, and it was unplaced, as the full Champions Day results show.

What the market charged on the day

The bookmakers' margin across the card averaged 17.7% over a fair book, which means they kept 15.0% of every pound staked at starting price. That figure was far from even across the day.

The cheapest race was the Champion Stakes, with 6 runners and an overround of 10.7%. The dearest was the Balmoral Handicap, a 20-runner cavalry charge down the straight mile, at 30.4%. The handicap cost 2.9 times as much as the feature race. A big field gives the bookmakers more runners to spread their margin across, and that is where it tends to be widest.

That is the uncomfortable part of the blind arms' best result. Fifth Column won in the race where a punter was paying the most for an opinion. One winner at 22/1 covers a lot of margin on one afternoon. Over a season, the margin is paid on every bet and the 22/1 winners are rare.

For one afternoon, then, the League's return of +47.0% left it clear of the margin it was paying. The best season record on the board is StableBet Rating on -7.2%. Every entrant's season figure is on the Silicon Tipster League, and the meeting itself has a hub at British Champions Day. How we freeze each festival's numbers, and what the record refuses to claim, is set out in how we measure a festival.

Champions Day tipster league FAQ

Did the AI tipsters make money at Champions Day?

On the day, 4 of the 14 entrants finished ahead, and two long-priced winners did almost all of that work. Over 7 races that is within what chance produces: a random pick in every race finished ahead in 17.8% of simulated trials. Each entrant's season record is the figure that matters, and on that record Stablebet Edge, the day's best performer, sits on -15.4%. Nothing here is evidence that an AI can beat the bookmakers.

Why did Stablebet Edge back Mount Atlas?

Its rule is to back the runner with the biggest gap between our model's estimate of its chance and the chance implied by the market. In the Long Distance Cup that runner was Mount Atlas, whom the model rated at more than twice the market's view. The rule picks a horse like this in every race, and most of them lose.

What is the difference between the blind and informed arms?

Every chatbot enters twice. The blind arm sees the form and no prices and is asked to name the winner. The informed arm sees the market as well and is asked where it disagrees with it. They are settled and reported separately, which is how the League measures whether seeing the prices helps a model or costs it.

Why does The Favourite differ from the starting-price favourite?

The Favourite backs the horse heading the morning market on the 09:05 card, which is fixed before racing. Prices move during the day, and at Ascot the morning favourite and the starting-price favourite were different horses in three of the races. We report both, labelled, because they can return very different figures over the same card.

Is any of this a tip?

No. Every pick described here was logged before the race and settled at starting price, win or lose, and the page reports what happened rather than what to back next time.

Bet responsibly. Nothing on this page is betting advice. If you bet, stake only what you can afford to lose, and if it stops being fun, stop. Help is at BeGambleAware.org. 18+.

Every figure here is pulled live from our data and nothing beats the bookmaker's margin. For whether anyone holds a real edge, see our track record. 18+, please bet responsibly.

More from Lab Notes