Averaging answers across several AI models works better than asking one model once — and a new study puts numbers on how much better.
Researchers ran 960 prompts across GPT-5, Gemini 2.5 Pro, and Claude Sonnet 4.5, testing eight estimation tasks. They tried two aggregation approaches: sampling the same model multiple times (intra-model) and combining outputs across different models (inter-model). Both cut error. The best strategies reduced mean absolute percentage error by up to 37 percentage points. The study also found that models showed signs of metacognitive awareness — their expressed uncertainty correlated meaningfully with how wrong they actually were (Spearman's rho between 0.242 and 0.568).
The finding matters because it gives organizations a practical lever they already control: run the same query against multiple models, or multiple times, and average the results. No new architecture required. The metacognition result is the more surprising piece — if a model's confidence intervals reliably track its error rate, that uncertainty signal becomes genuinely useful, not just decorative.
Human swarm intelligence research has shown crowds outperform individuals for decades; the open question was always whether the effect would survive replacing humans with AI. Apparently it does, at least on estimation tasks — though calling it "wisdom of the crowd" when the crowd is three API endpoints is doing some heavy lifting.