AI Agents Trail Human Average in 100-Plus Fantasy Soccer Tests as Forecasting Models Improve
Updated
Updated · WIRED · Oct 28
AI Agents Trail Human Average in 100-Plus Fantasy Soccer Tests as Forecasting Models Improve
3 articles · Updated · WIRED · Oct 28
Summary
A fantasy soccer team run by AI agents failed to beat the average human score over recent weeks, even after its manager expanded the system to more than 100 specialized bots.
Codex performed better once it was given a custom harness—agents split across news reading, results analysis, modeling and critique—but adding more agents mostly increased API costs rather than results.
Haifeng Xu of the University of Chicago said newer models are improving because they can absorb far more real-world information and coordinate more complex forecasting workflows.
In Xu’s tests, GPT-5 ranked best on prediction tasks ahead of Grok 4, Claude Sonnet, Gemini and Llama, though AI still often trails elite human superforecasters.
Benchmarks including Prophet Arena, Prediction Arena and the Metaculus Cup suggest that edge is narrowing, pointing to wider disruption for trading and other prediction-heavy work.