AI bots win the Metaculus Cup for the first time
Bots Laertes and ManticAI placed first and second among 885 forecasters, though a separate spring test of matched questions had the best humans still slightly more accurate.
- Benchmarks & progress
- Minor
Metaculus said bots finished first and second in its Summer 2026 Metaculus Cup, a quarterly tournament in which forecasters predict the outcome of real-world questions. A bot called Laertes, built by an individual developer, and a second bot, ManticAI, outscored all 885 human forecasters across 58 questions and roughly 88,500 individual forecasts, the first time an AI has topped the tournament outright. Bots are not eligible for prizes but compete on the same questions as human entrants.
Metaculus’s own announcement noted a structural point that qualifies the result: the Cup’s scoring rewards both accuracy and how consistently a forecaster updates a prediction as new information arrives, and a bot can update on every question continuously in a way no human forecaster can match, regardless of the two models’ relative skill at judging genuine uncertainty. A separate, controlled comparison run earlier in 2026 — matching ten of Metaculus’s top human forecasters against ten leading bots on 99 shared questions, over the same window — had found the humans slightly more accurate on average, a gap the tournament’s diligence-rewarding format does not capture.
The result adds to a small set of forecasting-tournament outcomes tracked as a proxy for how AI systems handle open-ended uncertainty about future events, distinct from the closed, verifiable problems most benchmarks test. Coverage of the win was consistently paired with the caveat about scoring mechanics, making this less a clean claim that AI forecasting judgement has overtaken the best humans than a demonstration that tireless, continuous updating is now enough to win a tournament scored the way Metaculus scores it.