How each AI's picks are scoring
Running points from a notional £10 per pick at SP. Tap a name to follow one AI.
Swipe the chart to see the full record →
Dashed: The Favourite, the bar to measure against. Whole record since 6 July.
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.
Which AI picks the most winners?
Five chatbots read each race twice: odds hidden to pick a winner, then odds shown to find a price the market got wrong. Every pick is logged before the off and settled at Starting Price.
- 01Choose a race below
- 02Compare picks, odds hidden
- 03See their value picks
- 04Check the scores after racing
Risked is 10 points a bet. The marker on each bar is break-even, every point risked won back. The records overlap within their ranges, so the order is a sort, not yet a ranking. The small line is the last fortnight.
Pick a race, see every pick
Choose a race from the timetable. Each AI's pick sits beside the horse's chance three ways: the market's view, our model's and, in the price job, the AI's own. Results appear once the race has settled.
19:40 Newcastle
Midnite: Built For 2026 Not 2006 Handicap · 14 runners · 5f · Standard
Who wins?
AI Picks the Winner · odds hiddenHouse entrants
Which price is wrong?
AI Finds the Value Bets · odds shownHouse entrants
Won by Kinetic Force. 0 of 5 named the winner with the odds hidden; 0 of 5 backed it with the odds shown.
Better at naming winners than at spotting wrong prices
After 136 days and 58,332 settled picks, the clearest thing the League has found is about the two jobs, not about which chatbot is cleverest.
With the odds hidden, every AI does better than when it is shown the odds and asked for a price the market got wrong. Claude is the best at naming winners, with a return of −4.3% (−1,949 points on a notional £10 a pick); hunting prices, the best any AI manages is −16.5%. For Claude the gap between the two records is wider than luck alone would explain.
Shown the market, they go after longer prices, so they win less often: 9–15% of their races, against 20–23% naming winners. That is the trade the price job asks for.
None of them is above break-even yet, and Claude's −4.3% is within luck of simply backing the favourite (−12.7%). The table below shows where each one stands today.
See the full breakdown ↗The League table
“Points” is each entrant's running total on a notional £10 a pick at Starting Price. “Return” is that total as a share of the points risked, so a return of zero is break-even. The form guide shows the last week of race days, gold for a day that finished ahead.
| # | Tipster | Bets | Won | Points | Return | Last 7 days | Fortnight |
|---|---|---|---|---|---|---|---|
| 1 | ClaudeAnthropic | 4,565 | 22% | −1,949 | −4.3% | ||
| 2 | GeminiGoogle | 4,614 | 20% | −4,897 | −10.6% | ||
| 3 | ChatGPTOpenAI | 4,623 | 23% | −5,129 | −11.1% | ||
| 4 | GrokxAI | 4,624 | 22% | −5,749 | −12.4% | ||
| 5 | DeepSeekDeepSeek | 4,502 | 21% | −7,334 | −16.3% | ||
| Benchmarks: our own rules, The Favourite is the bar | |||||||
| The FavouriteThe market | 4,569 | 29% | −5,818 | −12.7% | |||
| StableBet RatingStablebet | 1,167 | 18% | −1,584 | −13.6% | |||
Ranked by return over the whole record. Tap a row for that entrant's own page. Head to head, money and luck: the full breakdown ↗
How the League works
Every British and Irish race, each AI gets the racecard: runners, form, going, trip and class. It does two jobs. Who wins? The odds are hidden and it names the horse it thinks will win (AI Picks the Winner). Which price is wrong? It sees the market's chance for every runner, with the bookmaker's margin removed, and names the horse it thinks the market has priced too long, with its own estimate (AI Finds the Value Bets).
Picks are logged each morning and published before the first race, then settled at industry Starting Price after racing and scored in points, a notional £10 a pick, fallers and non-finishers counted as losers and non-runners void (0 points). 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. We call each chatbot through its provider's API (GPT-4.1, Claude Sonnet, Gemini, Grok and DeepSeek) and publish the exact model behind every pick: the board and the day files beside it are public. The raw board is settled at £1 a pick, so every points figure here is ten times its figure.
House entrants run on the same races: The Favourite (the morning favourite, the bar for Who wins?), the StableBet Rating (handicaps only), and our model twice in Which price is wrong?, as the Stablebet Model (its most likely winner) and Stablebet Edge (its biggest gap to the market). Each entrant has its own page: Claude · ChatGPT · Gemini · Grok · DeepSeek · The Favourite · StableBet Rating · Stablebet Model · Stablebet Edge.
Correction, 24 September 2026: The going was missing from our morning cards, 3 to 23 September. Read the note
Which AI is the best horse racing tipster?
Naming winners with the odds hidden, Claude has the best return so far, −4.3% over 4,565 bets at Starting Price, which is −1,949 points on a notional £10 a pick. The five AIs' ranges overlap, so none has pulled clear yet. The scoreboard updates after each day's racing.
What does the Tipster League measure?
Two jobs, on every British and Irish race. Who wins? hides the odds and asks each AI to name the winner from the form alone (the test we call AI Picks the Winner). Which price is wrong? shows the market's chance for every runner and asks for the horse the market has priced too long (AI Finds the Value Bets). Each job keeps its own record.
Why hide the odds from the AI?
Because the price already contains most of what is known about a race. An AI shown the odds can look sharp just by leaning on the favourite, which says nothing about its own reading. Hidden odds make it form a view of its own, and that view is what this test measures.
Should I follow ChatGPT's betting tips?
Use it as a quick read of a race, not as a tip. With the odds hidden, ChatGPT's picks win 23% of their races, close to the other AIs, and return −11.1% at Starting Price (−5,129 points on a notional £10 a pick). Nothing here is a signal to stake.
Do you bet real money on these picks?
No. Every pick is settled on paper at industry Starting Price and scored in points. 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. This is research, not a tipping service.
Research, not tips. Every pick is logged before the off and settled at industry Starting Price, wins and losses alike. Nothing here is a signal to stake. 18+ · If gambling is affecting you or someone you know, free and confidential support is at BeGambleAware.org.
