Which AI Picks the Most Horse Racing Winners? Inside the Proof Page

29 September 2026 | Updated 13 October 2026 | Realtime Tips

Ask any tipster for their record and you will get the good months. Ask the Advanced proof page and you get everything: 37,555 published selections from 10 models since 6 March 2026, of which 36,063 have settled for 10,452 winners (29.0%) and a combined -1,000.85 BSP LSP from 6 March 2026 to 12 October 2026. Nothing is removed once it is published, and the page is rebuilt every day as results settle.

Prices are BSP with SP fallback: level-stakes profit is calculated at Betfair Starting Price, with industry SP used only when no BSP exists, and Betfair commission is not deducted. Anyone can see every row up to yesterday. Signed-in Advanced members also see today's pending selections as they are published, so the members' view is the same record one day ahead.

Every figure here is a model's record on all of its picks. No filter has been fitted to the results, so nothing in these tables was chosen with hindsight. This article walks through the page, puts all 10 models side by side on their complete records and gives an honest early report on the newest models (Grok 4.7, GPT-6 Astra, Claude Fable 5.1).

View the proof page ›

How to read the proof page in 60 seconds

• The overview table has one row per model with its complete record: every selection the model published, settled at level stakes. Nothing is filtered out, so the losing models are there alongside the rest.
• Open a model by clicking its name to see its record on its own, including a table of its selections by race type.
• The daily grid colours each day green or red by that day's BSP LSP. Click a day to open a dialog listing every selection with its price, result and confidence, so a good or bad day can be checked runner by runner.
• CSV and XLSX exports give you the same rows the page uses, so you can rebuild any figure on it yourself.
• Price basis: everything labelled BSP LSP is level stakes at Betfair Starting Price with industry SP as the fallback. Commission is not deducted.

The league table: all picks

This is every model on its complete record, sorted by BSP LSP. "Days" is the number of days on which the model published selections, and "Avg price" is the average decimal proof price of its settled selections (BSP with SP fallback), which tells you whether a strike rate came from short-priced favourites or bigger prices.

2 of the 10 models are in profit at level stakes on their complete record, led by Grok 4.5 at +51.00, while GPT-5.5 sits at the bottom on -282.39.

The biggest body of evidence belongs to GPT-5.4 with 8,090 settled selections. Among models with at least 200 settled, the best return on stakes is Grok 4.5 at 1.7%, which is the fairest way to compare records of very different lengths. The highest strike rate in that group is Claude Fable 5.1 at 31.0%. The highest BSP LSP figure is Grok 4.5 at +51.00.

The newcomers' report card

A model appears on the proof page from its first published day, not from the first day it looks good. These are the models with fewer than 45 days of published selections, newest first.

Grok 4.7 (current xAI model)

First published on 22 September 2026, so 21 days of published selections: 240 winners from 837 settled (28.7%), -43.10 BSP LSP and 26 pending. By race type it is best over Hurdle (+12.84 from 162) and weakest over Flat (-59.39 from 578). That is still a small sample, so treat the shape of the record as provisional. Open the Grok 4.7 record to watch it grow day by day.

GPT-6 Astra (current OpenAI model)

First published on 5 September 2026, so 37 days of published selections: 425 winners from 1,410 settled (30.1%), -63.52 BSP LSP and 48 pending. By race type it is best over N H Flat (+0.66 from 26) and weakest over Flat (-50.00 from 1,050). That is a usable sample, although still short of a full season. Open the GPT-6 Astra record to watch it grow day by day.

Claude Fable 5.1 (current Anthropic model)

First published on 2 September 2026, so 41 days of published selections: 490 winners from 1,583 settled (31.0%), +17.45 BSP LSP and 48 pending. By race type it is best over Flat (+23.55 from 1,182) and weakest over Hurdle (-10.38 from 252). That is a usable sample, although still short of a full season. Open the Claude Fable 5.1 record to watch it grow day by day.

Recent form: the last 30 days

Each row covers the 30 calendar days ending on that model's most recent settled day, using all selections. For the newest models this is simply their whole record so far.

2 models are in profit over the latest 30 days. Claude Fable 5 has the best recent figure at +28.63 from 1,147 settled, and GPT-5.4 the worst at -106.30 from 1,133. A 30-day window is exactly the kind of sample that flatters or condemns a model unfairly, so use it to spot a change of direction, not to pick a winner.

Three mistakes people make with this page

• Treating strike rate as profit. Among models with at least 200 settled, the highest strike rate is Claude Fable 5.1's 31.0% winners from 1,583 settled selections since 2 September 2026. A model with a high strike rate can still lose money if the prices are short, and a model with a lower strike rate can be the better bet at bigger prices. Read win rate, average price and BSP LSP together.
• Reading a tiny sample. A new model's first week can look like genius or disaster. Under 300 settled selections tells you almost nothing; a few thousand starts to mean something. The Settled column is there for a reason.
• Comparing totals over different windows. A model that started in March has had months longer to build up a profit or a loss than one that started last month. Check the First day and Settled columns, and compare ROI rather than the BSP LSP total when the records are of very different lengths.

What we are watching next

• The newcomers' samples. Grok 4.7, GPT-6 Astra, Claude Fable 5.1 have fewer than 45 days of history. Their rows on the proof page grow every racing day, and the sample-size warnings above apply until they reach a few thousand settled selections.
• The jumps season. From October the mix shifts towards Chase and Hurdle races, which several models have handled differently from the Flat. The race-type tables on each model page are where that will show first.

None of that is a promise of profit. It is a promise that the record stays complete and public, which is the only basis on which any of these models deserves your attention.

See the full record

Racing predictions are uncertain. Only bet what you can afford to lose and read our responsible gambling guidance.