The AI handicapping contest
What can be verified about the 2025 contest, what cannot, and where the 2026 numbers stand. Everything on this page is labelled by how well it is sourced.
The table below is not our own reconstruction. It is the leaderboard as the Internet Archive captured it on 15 August 2025, mid-contest, from this site. It is a third-party, timestamped record.
Leaderboard as of 15 August 2025
| Model | Record | Units | Win % | Wagered | ROI |
|---|---|---|---|---|---|
| Claude | 58-38-1 | +32.62u | 60.4% | 213.25u | +15.3% |
| Gemini | 59-39-1 | +22.08u | 60.2% | 178.14u | +12.4% |
| ChatGPT | 67-41-1 | +20.55u | 62.0% | 207.75u | +9.9% |
Picks published on 9 August 2025
Published the morning of the games, before results, as the archive shows.
| Model | Card that day |
|---|---|
| Claude | 3u Royals ML (-116) · 2u Astros ML (-122) · 2u Mets ML (+118) · 1u Nationals ML (+155) · 1u Phillies ML (+120) |
| ChatGPT | 3u Astros ML (-122) · 3u Guardians ML (-132) · 2u Giants ML (-161) · 2u Cubs ML (-108) · 1u Rays F5 ML (+111) · 1u D-backs -1.5 (-110) |
| Gemini | Astros ML (-122) · Giants ML (-161) · Guardians ML (-132) · Astros TT Over 4.5 (+114) · COL/ARI Over 9.5 (-110) · TOR/LAD Over 8.5 (-119) |
This site previously published final 2025 totals for each model. Those finals cannot be reproduced. We searched for a source ledger across local storage, the web server and every other site on it, the GitHub account and its code search, prior working copies, the site's own files and the Internet Archive. No per-pick record exists for the final week of the contest, or for any week. Rather than restate numbers we cannot stand behind, they have been removed. The two archived captures above are what survives, and they are shown exactly as captured.
2026 contest
The 2026 material is a different case: its arithmetic checks out. The tracker's five model records sum exactly to its stated 469 graded picks, and the diagnostic's four sum exactly to its stated 314. The two reconcile with each other as successive snapshots. Both are kept, each stamped with the date its data stops.
2026 tracker
Five models, 469 graded picks, combined 210-232-27. Data ends 19 April 2026.
2026 diagnostic
Four models, 314 graded picks, combined 138-155-21 and −95.08 units. Data ends 10 April 2026.