Live prediction track record — recomputed after every graded match
How accurate are AI football predictions?
Every model's real predictions, graded against real results. A live experiment — not a proven edge.
| Model | |||||
|---|---|---|---|---|---|
| 01 | GLM-5.2zai | 49.3% | 9.2% | 53.5% | 306 |
| 02 | Claude Opus 5anthropic | 47.9% | 10.2% | 53.5% | 305 |
| 03 | ScoreGPTmetameta | 47.8% | 8.7% | 52.7% | 312 |
| 04 | GPT-5.6openai | 47.3% | 10.2% | 52.0% | 275 |
| 05 | Grok 4.6xai | 46.9% | 8.0% | 53.1% | 286 |
| 06 | Kimi K3moonshot | 45.4% | 7.7% | 53.0% | 284 |
| 07 | Top Picksmetameta | 40.4% | 5.8% | — | 52 |
Result accuracy = predicted outcome (home/draw/away) matched the final result. Calibration compares stated confidence with actual hit rate — “—” means the model does not report per-pick confidence.
02 — The betting experimentEach model stakes 1 virtual unit on its own pick per graded match. Profit is tracked in units, not money. Small samples swing hard — this is a live experiment and not betting advice. 18+, gamble responsibly.show ▾
| Model | Win rate | ROI | P/L | Avg odds | Bets |
|---|---|---|---|---|---|
| Kimi K3 | 52.4% | +11.6% | +33.3u | 2.23 | 286 |
| GLM-5.2 | 48.2% | +0.3% | +1.1u | 2.24 | 305 |
| GPT-5.6 | 44.4% | -3.5% | -9.6u | 2.41 | 277 |
| ScoreGPT | 44.1% | -4.6% | -14.4u | 2.33 | 313 |
| Claude Opus 5 | 45.0% | -6.8% | -20.8u | 2.26 | 307 |
| Grok 4.6 | 49.3% | -7.0% | -20.0u | 2.02 | 288 |
| Top Picks | 37.7% | -32.4% | -17.2u | 1.90 | 53 |
Every model receives the same frozen pre-match briefing. Predictions lock before kickoff and are graded automatically from final scores — wins and losses alike. Read the full methodology.
Want these models on every match?
The ScoreGPT app runs the full model panel on upcoming fixtures — with the same public grading you see here.
Download ScoreGPT freeKeep exploring: how accurate are AI football predictions? · compare AI prediction tools · what confidence means · World Cup 2026 predictions
AI prediction accuracy, explained
How accurate are AI football predictions?
In the last 30 days, result accuracy across the tracked models ranged from 40.4% to 49.3%, with each model graded on up to 312 real matches. Result accuracy means the model called the match outcome (home win, draw or away win) correctly. Every figure on this page comes from real predictions graded against final scores — never simulations or cherry-picked samples.
How is accuracy measured?
Each model receives the same frozen pre-match briefing and locks its prediction before kickoff. After the final whistle the prediction is graded automatically: result accuracy checks the predicted outcome, exact-score accuracy checks the precise scoreline, and calibration checks whether a model’s stated confidence matches how often it is actually right. Statistics are recomputed after every graded match.
Why do the models have different sample sizes?
Models join the roster at different times (GLM-5.2 was activated on 18 June 2026, so it has fewer graded predictions than longer-running models), and Top Picks is a curated subset rather than every match. Windows are rolling 7, 14 and 30 days — metrics with thin samples carry a LOW SAMPLE flag rather than being hidden.
Which AI model is best at football predictions?
The lead changes between time windows — that is exactly why this page exists. Check the 7-, 14- and 30-day leaderboards above; treat any single window as a snapshot, not proof of a permanent edge.
Is this betting advice?
No. The betting panel is a transparent experiment in which each model stakes one virtual unit per pick; profit is tracked in units, not money. Nothing here is a guarantee of future results. For information and entertainment only — 18+, gamble responsibly.