Enterprises using multiple AI models are underestimating failure rates by 2.25x

A study of 67 frontier AI models from 21 providers finds that enterprises often overestimate the reliability gains from routing queries across multiple models. The key issue is the co-failure ceiling: the share of prompts where every model in a pool fails at the same time, limiting the upside of orchestration.

The findings challenge the assumption that more model diversity automatically produces safer or better AI systems. As enterprises add routers, cascades, and agent mixtures, they may incur higher latency, infrastructure complexity, and governance risk without meaningful performance gains unless they first measure shared failures.