How Claude Picks Horses: The Careful Punter

Hand Claude a bare racecard, the prices stripped out, and it reads a race the way a careful punter does: trainer first, then the form, then who is riding.

Across 3,798 settled picks with the odds hidden, its flat-stake return landed at −8.3%. That is a loss, as every arm in the league is, but a shallow one, and the confidence range of −16.0% to +0.1% crosses zero: this sample cannot separate a small loss from a small gain.

Before that tempts anyone, read the small print in the same breath. That range sits well inside the swing ordinary variation hands out for free, so treat it as a curiosity, not a signal, and certainly not a reason to follow anything blind. The honest, and more interesting, story is not the size of the number. It is how Claude arrives at a pick, and how sharply that changes the moment you let it see the odds.

This is one model's fingerprint from the Silicon Tipster League, where five chatbots tip real British and Irish races every day in two modes, blind and informed. For where it sits among the rest, here is the full verdict on all five AIs.

What the careful read looks like

Ask Claude to justify a blind pick and the same words keep surfacing. Its three most-used reasons are trainer, form and jockey, in that order. It leads with the yard and the recent form, the way a paddock regular does, rather than reaching for a hunch.

Does that make it a favourite-backer? Sort of, but choosier than you might guess. With the odds hidden, all five AIs drift towards the shorter prices, and Claude is no exception: the typical price it lands on is around 6.23, and its median pick sits at market rank 3 by price, which is where all five land blind. But it backs the outright favourite only 26% of the time. Most of its bets live in the mid-priced middle of the card, 54% of them, with a healthy 31% out among the longer prices. It leans short without hugging the chalk.

Line those picks up against every simple staking rule we test and the closest match is "back short-priced favourites", though the resemblance is only moderate (0.38 on a nought-to-one scale). It wins 21.5% of its blind races. The market brings home 37.7% on the same cards, a standing reminder of how good the crowd already is at this.

Show it the market, and it turns hunter

Give Claude the same races, let it see the odds, and ask it to find value rather than the winner, and it becomes a different animal. Its reasons reorder around a single word: the top theme flips to value, ahead of form and trainer. The careful favourite-leaner is gone.

It goes straight to the outsiders. Shown the market, the share of its picks that are the favourite falls to 0%, while 70% of them sit out among the longest prices, at a typical price of about 11.10. Its closest strategy match swings all the way from short-priced favourites to "back the outsider", and this time it barely resembles any rule we test (0.04 on the same nought-to-one scale). Put plainly, when it hunts value it is largely freelancing, not following a template.

So the behaviour change is real, it is large, and it is measurable. That is the finding, and the price of it is now measurable too. The value return reads −24.6% over 3,441 settled picks, and the range around it, −34.4% to −14.3%, no longer crosses zero. When this was first written the arm was weeks old and that number could have been luck in either direction. It is not any more: on the sample it now holds, this is the heaviest loss of the five value arms, and swinging from careful favourite-leaner to outright contrarian is what it cost. A model that changes its mind this completely when you show it a price is worth understanding. It is not worth following. Research, not tips.

How to use the careful one well

Treat a Claude tip as a fast, competent second read of a card, not a tipsheet. Its strengths are real even where its edge is not: blind, it hands you a sober, form-led shortlist built on trainers and recent runs; shown the odds, it turns into a contrarian radar pointing at prices the favourite-backers walk past.

Three things to hold in mind. Read its reasons, not just its pick, because the trainer-and-form logic is the useful part. Remember it shares our own biases: it leans to the short prices blind and swings hard the other way once it sees the market, the two traps human punters fall into most. And take every figure here as a snapshot of a young, noisy experiment, not a system that makes money.

You can watch it keep its own score. Here is its live record, the blind board it plays in at AI Picks the Winner, and its value picks at AI Finds the Value Bets. For how it compares with the other four, read the full verdict on all five AIs.

Bet responsibly. Nothing on this page is a tip or 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+.