AI wins the Metaculus forecasting cup for the first time: the top human only places 3rd

Metaculus Cup is a seasonal tournament where people and bots answer the same questions in one overall ranking. In summer 2026, a machine won it for the first time.

Humans used to win. In FutureEval's controlled comparison in spring 2026, ten Metaculus pros beat the ten best bots by 1.25 points per question across 99 shared questions. The margin was within the margin of error, but 9 out of 10 pros individually outperformed any single bot.

The results from 18.09 reversed the picture. Laertes and ManticAI took 1st and 2nd place, FutureSearch came 5th, and the tournament's best humans remained 3rd and 4th. The Cup drew 885 participants and 88,558 forecasts across 58 questions.

The difference is not the model itself but the system built around it. In FutureEval's spring measurement across 297 questions, the top bots scored average marks of 18.90, 18.25, and 16.87, while the best unaugmented LLM, GPT-5.1, scored 11.32.

A professional forecaster's report costs more than $10,000 and takes a week to prepare. A FutureSearch forecast takes minutes and a few dollars: Scott Alexander calculated about $8 per question, and the initial $20 in credits covers roughly 4 questions.

Build your own bot in 5 minutes. Fork the Metaculus/metac-bot-template repository, put two keys (METACULUS_TOKEN and OPENROUTER_API_KEY) in GitHub Secrets, and enable Actions. The workflow then answers new tournament questions every 20 minutes, while it picks up Metaculus Cup questions once every 2 days. The token is issued after registration on the FutureEval participate page, while free OpenRouter credits are available through the form in the template README.

Bots are gaining about 0.9 Metaculus Elo points per month, according to Scott Alexander's estimate.