Metaculus's Spring 2026 FutureEval finds AI forecasting bots nearly match top human forecasters, gap not statistically significant

Metaculus ran its Spring 2026 FutureEval tournament (7 Jan to 15 Apr 2026): 173 forecasting bots and a 10-person Metaculus Pro human team were scored on overlapping questions, with 297 questions scored overall and 99 used for the head-to-head comparison. The Pro team beat the top-10 bot team by 1.25 points on average, but this gap is not statistically significant (p=0.247). That is a sharp narrowing from Q2 2025, when bots lost by roughly 20 points. Top bots mostly ran on GPT-5.2 or newer, with 9 of 10 surveyed top performers reporting GPT-5.4, and agentic/scaffolded bots beat simple single-prompt setups.

This gives builders of automated forecasting or research-triage systems a concrete, citable data point: current-generation models with agentic scaffolding are now statistically indistinguishable from skilled human forecasters on Metaculus-style questions, at least in this one tournament. That is useful for anyone deciding whether to trust or deploy bot-based forecasting pipelines, and it points at GPT-5.2+ plus scaffolding (rather than single-prompt calls) as the practical recipe. Impact is limited to forecasting and prediction-adjacent work; it says nothing about broader reasoning, real-world decision-making, or high-stakes domains outside structured question tournaments.

The headline framing ("gap is nearly gone") oversells a null result: p=0.247 means the study cannot distinguish bots from humans, not that bots have proven parity, and the 9-of-10-GPT-5.4 model-attribution claim comes from a voluntary self-reported survey, which is also non-significant and prone to selection bias. Metaculus is also grading its own product/tournament, so some house-brand optimism should be discounted even though the write-up is unusually transparent about these caveats.

1. 01Whether the Pro-vs-bot gap flips to bots winning (or stays statistically insignificant) in the next FutureEval season, ideally with a larger overlapping-question sample to firm up significance
2. 02Independent replications or third-party forecasting benchmarks (e.g. from other platforms) testing GPT-5.4-class agentic bots against human forecasters outside Metaculus's own tournament structure
3. 03Whether any organisation reports actually replacing or augmenting human forecasting/analyst teams with bot ensembles based on this result, rather than just citing it as evidence