AI football predictions
AI vs ChatGPT vs bookmakers: who predicts football best?
9 predictors — large language models and a statistical model — called the same football matches before kick-off: 553 graded fixtures from 12 Jul 2026 to 18 Oct 2026. We compared every one of them with the bookmaker favourite on exactly the same games.
Numbers recalculated 19 Oct 2026 from graded fixtures
Short answer
The scoreboard
Every model sees the same fixture list and must commit to one outcome — home win, draw or away win — before the match starts. Because models joined the panel on different dates, the fair comparison is the last three columns: the model and the bookmaker favourite on the same matches.
| Model | Graded | Hit rate | Same fixtures | Model | Bookmaker fav. | Difference |
|---|---|---|---|---|---|---|
| Claude Haiku 4.5 (retired)Anthropic | 384 | 53% | 360 | 53% | 52% | +0.8 pts |
| Gemini 2.5 FlashGoogle | 414 | 52% | 390 | 52% | 52% | +0.3 pts |
| GPT-5.6 SolOpenAI | 413 | 53% | 393 | 53% | 53% | 0.0 pts |
| Qwen 2.5 32B (retired)Alibaba (local) | 406 | 52% | 382 | 53% | 53% | 0.0 pts |
| Kimi K3Moonshot | 407 | 54% | 386 | 55% | 55% | -0.3 pts |
| Scorebase ModelElo + Dixon-Coles + market blend | 532 | 51% | 508 | 52% | 52% | -0.4 pts |
| Grok 4.1 Fast (retired)xAI | 375 | 51% | 351 | 52% | 53% | -1.1 pts |
| Grok 4.7xAI | 25 | 64% | 25 | 64% | 60% | small sample |
| Claude Fable 5.1Anthropic | 23 | 61% | 23 | 61% | 61% | small sample |
"Bookmaker fav." = the outcome with the highest bookmaker-implied probability after removing the margin. Rows with fewer than 30 shared fixtures are not ranked. Predictions created after kick-off are excluded (56 rows seeded on 2026-06-27, the first day of operation).
Why football is so hard to call
In these 553 matches the home side won 44% of the time and 27% ended in a draw. A draw is rarely the single most likely outcome, so almost every model — and the bookmakers — loses those games by design. That is why a football match-winner hit rate in the 50s is normal, and why anyone advertising 80–90% on single football picks deserves scepticism.
When the models agree
On 405 matches at least three of the current models made a prediction and all picked the same outcome. Those unanimous picks were right 53% of the time, against a typical single-model hit rate of 52%. Agreement on its own adds little — in 96% of those games the shared pick was also the bookmaker favourite.
What the three kinds of predictor actually do
- Large language models (GPT, Claude, Grok, Gemini, Kimi) predict from the match context in their prompt and must commit to one outcome before kick-off.
- The statistical model rates every team with Elo and a goal model, then blends in market odds. Its job is calibrated probabilities, not bold calls.
- Bookmaker odds aggregate the money and information of the whole market. They are the benchmark every predictor has to beat.
How this was measured
Predictions are stored with a timestamp and only counted if they were made before kick-off. Postponed and abandoned matches are not graded, and nothing is deleted or re-scored afterwards. The full leaderboard across all markets and sports is on the accuracy page; the model-by-model scorecard is published by Scorebase, which runs the models.
These are statistical estimates, not betting advice. Any single match can go against every prediction.