Study Shows Combining AI Models Has a Hard Accuracy Ceiling, Not a Fix for Errors

A new paper (arXiv 2606.27288) analyzing 67 frontier language models from 21 providers finds that ensemble strategies — routing, voting, or mixture-of-agents — cannot exceed an accuracy of 1 minus the rate at which all models fail on the same question simultaneously. This shared co-failure rate, termed β, is the true limiting factor, and no combination method can recover a correct answer if none of the individual models produced one. The research also reveals that pairwise error correlation, the standard metric used to assess model diversity, systematically underestimates β and cannot detect higher-order co-failure patterns. Notably, on a hard science benchmark, co-failure nearly vanished under multiple-choice format but jumped to β = 0.127 when the same questions were posed as free-response, suggesting that much of the apparent diversity among frontier models is an artifact of multiple-choice scaffolding. The findings challenge a widely held engineering assumption that redundancy in AI systems reliably improves reliability.