A new study evaluating 67 frontier models from 21 providers finds that enterprises routing queries across multiple AI models assume each covers the others' blind spots, but mathematically that assumption is flawed, leading to underestimated failure rates.