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

Enterprises deploying multiple AI models to cover each other's weaknesses are experiencing failure rates 2.25 times higher than they expect. A study of 67 frontier models from 21 providers reveals why: companies ignore what researchers call the "co-failure ceiling."

The typical assumption behind multi-model orchestration is straightforward. Route a query to a coding specialist, a logic specialist, and a generalist model. When they disagree on outputs, at least one should be right. The failures won't overlap.

That logic fails mathematically. The actual constraint on orchestration depends not on disagreement rates, but on the percentage of prompts where every model in the ensemble generates wrong answers simultaneously. That's the co-failure ceiling.

When enterprises build routing infrastructure around model disagreement, they're optimizing for the wrong metric. They assume diversity of architectures and training data prevents correlated failures. Instead, many frontier models share underlying vulnerabilities. A prompt that stumps one tends to stump others, even if their wrong answers differ.

The study tested this across coding, logic, and general knowledge tasks. Models failed together far more often than enterprise teams predicted. This gap between expected and actual failure rates compounds across production systems handling thousands of queries daily.

The implications are direct. Enterprises spend resources on sophisticated routing logic and monitoring dashboards that don't address the core problem. When the co-failure ceiling is low on a particular task class, no amount of orchestration between models in that pool improves outcomes. Adding a fourth model doesn't help if all four stumble on the same types of prompts.

The finding matters for critical applications. Medical diagnosis systems, code review tools, and financial analysis platforms rely on ensemble confidence to justify automation. If the co-failure ceiling is genuinely 2.25x lower than expected, those systems are riskier than teams believe.

The solution isn't necessarily to abandon