Words of estimative probability (WEPs), such as ‘possible’ and ‘a good chance’, provide an efficient means for expressing probability under uncertainty. Current semantic theories assume that WEPs denote crisp thresholds on the probability scale, but experimental data indicate that their use is characterised by gradience and focality. Here, we implement and compare computational models of the use of WEPs to explain novel production data. We find that, among models incorporating cognitive limitations and assumptions about goal-directed speech, a model that implements a threshold-based semantics explains the data equally well as a model that semantically encodes patterns of gradience and focality. We further validate the model by distinguishing between participants with more or fewer autistic traits, as measured with the Autism Spectrum Quotient test. These traits include communicative difficulties. We show that these difficulties are reflected in the rationality parameter of the model, which modulates the probability that the speaker selects the pragmatically optimal message.