The Rational Speech Act (RSA) model has been employed to explain word learning and inferences based on the costs of forms. Here, we focus on a hitherto untested and counterintuitive cost-based effect predicted by RSA: In learning a lexicon with two forms and meanings, learners should prefer an ambiguous costly form over an ambiguous cheap form. We demonstrate this prediction in an RSA model. We then measure reaction times and lexicon ratings in a novel word learning task to test whether a lexicon with an ambiguous costly form is less surprising than one with an ambiguous cheap form. We found no clear evidence for this effect in either measure. We discuss alternative explanations for documented cost-based effects, and the possibility that cost-based inferences may not occur during learning.