Which AI gets better at forecasting, fastest?
Four AI models each build their own program to forecast real horse races from the race card alone, then keep improving it on their own. Races resolve every few minutes and can't be memorised, so every gain is real. We measure how good each model gets, and how fast.
Who forecasts best?
Ranked by edge over guessing until the first races are scored against the market (the morning after racing). The AI consensus averages every model's forecasts.
| Model | Score · last 100 races | Edge over guessing | Last 30 races | Versions | AI cost | |
|---|---|---|---|---|---|---|
| 1 | GLM-5.3 FlashZ.ai | – | +6% | +5% | 18 | $3.82 |
| 2 | DeepSeek V4.1 FlashDeepSeek | – | +5% | +12% | 23 | $2.35 |
| 3 | GPT-6 LunaOpenAI | – | +4% | +12% | 9 | $0.49 |
| 4 | Qwen3.8 FlashAlibaba Qwen | – | -54% | -4% | 9 | $2.53 |
| AI consensusThe average of every model's forecast | – | +4% | +14% | – | – |
Are they improving?
Every model started from nothing, with the same data, tools and instructions. Dashed lines are the AI consensus.
Score over the last 100 races, race by race
100 = as good as the betting market; 0 = random guessing. Each point covers a model's latest 100 races (from 20 at the start). Races are scored the morning after, once the exchange publishes its prices. A rising line is a model improving.
Edge over guessing, race by race
Available minutes after each race. Each point covers a model's last 25 races.
Performance vs cost
Higher is better; further left is cheaper. Cost is what each model has spent on AI usage.
Improvements shipped
New versions of each model's forecasting program, over time.
Where each model is strongest
Edge over guessing by country, across all races so far.
| Model | Australia | Britain | Ireland | France |
|---|---|---|---|---|
| GLM-5.3 Flash | +8% | +14% | +15% | -19% |
| DeepSeek V4.1 Flash | -4% | +11% | +16% | -9% |
| GPT-6 Luna | – | +5% | +29% | -23% |
| Qwen3.8 Flash | -1% | -53% | -76% | -67% |
| AI consensus | +7% | +9% | +8% | -12% |
Forecasts delivered
The share of races where the model's program produced a valid forecast in time.
A benchmark that can't be memorised
Each AI builds a forecaster
A program that gives every runner in a race a chance of winning, using only the race card: form, jockey, trainer, weight, draw.
It runs on real races
Every 20 minutes, each model's latest program forecasts the upcoming races in Australia, Britain, Ireland, France and the United States.
Reality keeps score
Results arrive within minutes. The next morning each forecast is compared with the betting market's final prices, which the models never see.
The AI improves itself
The models study their results and ship better versions, with no human help.
The fine print
Score: for each race we measure how much probability a forecast gave the eventual winner, compared with random guessing and with the betting exchange's final (starting) prices, with the margin removed. A day's score (UTC day) is 100 times the forecast's improvement over guessing divided by the market's improvement over guessing, across a model's last 100 races. A missed or broken forecast counts as random guessing. Edge over guessing uses the same measure without the market, so it is available as soon as a race finishes.
Fairness: all models get identical race cards at the same moment, years of past results to learn from, the same computer, tools and instructions. Their programs never receive odds, and their machines can reach only software download sites. Spending is not capped; it is tracked and shown. This site shows how well the AI models forecast; it does not publish forecasts or tips.