— Monthly edition
AI prediction accuracy report: September 2026
10 AI models, 1,469 games, 14,596 graded picks (7 Sep – 30 Sep 2026).
Every pick is made before the game and graded after it. Wins and losses both count.
Key findings
- Claude Sonnet 5 picked the right winner most often: 55.2% of 1,469 picks. Mistral Medium 3.5 was last at 52.8%, so first to last was only 2.4 points.
- When every model agreed (757 of 1,469 games), the pick was right 64.1% of the time, against 55.3% for the majority pick across all games.
- Confidence means something: picks made with 70% or more confidence were right 80.5%, picks under 50% only 37.6%.
- Every model backed the home team more often than home teams won: 55.1% to 70.4% of picks, against 50.6% of games won at home.
- In football (soccer), where draws make it a three-way call, the majority of the models was right 47.8% of 592 matches, against 43.4% for always backing the home team and 33.3% for a random pick.
Accuracy by AI model
Share of picks where the model called the winner (or the draw) correctly, all sports.
| # | Model | Right winner | Picks | Exact score | Backed home team |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5 |
55.2%
|
1,469 | 6.3% | 70.4% |
| 2 | DeepSeek V4 Pro |
54.8%
|
1,469 | 6.3% | 63.7% |
| 3 | GLM 5.3 |
54.6%
|
1,422 | 6.5% | 66.9% |
| 4 | Grok 4.6 |
54.5%
|
1,469 | 7.1% | 66.8% |
| 5 | Kimi K3 |
54.3%
|
1,422 | 6.9% | 59.2% |
| 6 | GPT-5.6 Luna |
54.1%
|
1,469 | 6.7% | 58.1% |
| 7 | Gemini 3.8 Flash |
54%
|
1,469 | 8% | 59.7% |
| 8 | Qwen 3.8 Max |
53%
|
1,469 | 6.2% | 58.8% |
| 9 | MiMo V2.5 Pro |
52.8%
|
1,469 | 6.1% | 67.9% |
| 10 | Mistral Medium 3.5 |
52.8%
|
1,469 | 5.4% | 55.1% |
When the models agree
- All models agreed (757 games)
- 64.1%
- Majority pick, all games
- 55.3%
- Best single model (Claude Sonnet 5)
- 55.2%
- Home team won
- 50.6%
Does AI confidence mean anything?
- Stated confidence under 50% (1,483 picks)
- 37.6% right
- Stated confidence 50–59% (7,680 picks)
- 50.8% right
- Stated confidence 60–69% (4,788 picks)
- 60.6% right
- Stated confidence 70% and up (645 picks)
- 80.5% right
Accuracy by sport
Sports with fewer than 30 graded games are left out of this table but count in the totals.
| Sport | Games | Majority right | All agreed, right | Home won | Best model |
|---|---|---|---|---|---|
| Football (soccer) | 592 | 47.8% | 62.6% (227) | 43.4% | DeepSeek V4 Pro · 49.3% |
| Ice hockey | 351 | 59.9% | 61.7% (206) | 56.7% | Grok 4.6 · 61.5% |
| Baseball | 339 | 55% | 59.6% (198) | 53.1% | Gemini 3.8 Flash · 56% |
| Volleyball | 57 | 75.9% | 78.9% (38) | 54.4% | Kimi K3 · 80.7% |
| Handball | 48 | 68.1% | 80.6% (31) | 60.4% | Gemini 3.8 Flash · 75% |
| NFL | 48 | 60.9% | 70% (30) | 56.3% | Grok 4.6 · 68.8% |
Use the data
Free to republish under CC BY 4.0. Please credit TuringStats with a link to this page.
Cite as: TuringStats AI Prediction Accuracy Report (September 2026), turingstats.com/ai-accuracy-report
Download every graded pick (CSV)Embed the live leaderboard
Paste this into any article or blog. It updates on its own.
Method
- All 10 models get the same match data for every game (form, venue, head-to-head and, where available, team news), never the betting odds. When the prompt changes, it changes for every model at once.
- Each model returns a winner (or a draw where the sport allows it), a score and a confidence from 0 to 100. Only its latest pick before the game counts.
- A pick is right when the result matches. "Majority" is the result most models picked (ties are left out); "all agreed" means every model picked the same result.
- Results come from official match data. The line-up has been fixed since 7 September 2026; earlier picks are excluded.
FAQ
Which AI is best at predicting sports?
Claude Sonnet 5 picked the right winner most often: 55.2% of 1,469 picks. Mistral Medium 3.5 was last at 52.8%, so first to last was only 2.4 points.
Are AI sports predictions accurate?
Better than chance, far from certain. Across 1,469 games the majority pick of the models was right 55.3% of the time, and 64.1% when every model agreed.
How is the accuracy measured?
Every model predicts the winner (or a draw), the score and its confidence before kick-off. After the final whistle each pick is graded right or wrong. Only the latest pick before the game counts, and every graded pick is in the CSV download.
Can I use these numbers?
Yes. The report and data are free to use under CC BY 4.0: credit "TuringStats" with a link to this page.