A team routing queries across a coding specialist, a logic specialist, and a generalist model assumes each will cover the others' blind spots. A new study evaluating 67 frontier models from 21 providers shows that assumption is mathematically flawed — and the flaw has a name: the co-failure ceiling.
The assumption works like this: as long as two models don't usually fail on the exact same prompts, combining them is supposed to create a safety net against failures. In reality, the real limit on orchestration is not how often models disagree, but the percentage of prompts where every model in the pool gives the wrong answer at once. By ignoring the co-failure ceiling, enterprises are building complex, expensive routing infrastructure to chase performance gains that do not exist.
To orchestrate multiple language models, developers typically rely on three architectures: model routers that act as traffic cops, sending complex queries to expensive models and simple queries to cheaper ones; and cascades that send every prompt first to cheaper models. The hidden costs of these strategies multiply when the co-failure ceiling kicks in, as the redundancy designed to catch errors often fails simultaneously across models.
Fortunately, developers can use this same math to build a cost-free test that determines exactly when multi-model orchestration will actually pay off. The study suggests that without such a test, enterprises are systematically overestimating the reliability of their multi-model stacks and wasting resources on redundant inference compute.
The findings challenge a core tenet of AI infrastructure strategy — that more models necessarily yield more resilience. Instead, the data indicates that careful model selection and failure analysis matter more than raw orchestration complexity.