Numerical noun phrases (NNPs) are ambiguous between at least two and exactly two interpretations. This ambiguity has been commonly discussed as a case of Scalar Implicature (SI), where meaning is enriched by exclusion of an alternative. But the SI approach to NNPs has also been widely challenged. We tested NNPs and scalar expressions in inference tasks and found that scalars were sensitive to a manipulation that altered the relevance of alternatives, whereas exactly readings for NNPs were not. Our findings provide a theory-critical challenge to the SI view of NNPs and support alternative views.