During human communication, speakers balance informativeness and effort by tailoring their language to their audience. Large Language Models (LLMs) appear human-like in their communication and succeed at some tasks thought to involve social reasoning about interlocutors (in humans). Here, we tested audience design in LLMs using tasks modeled on (Isaacs & Clark, 1987). In Experiment 1, replicating findings with humans, LLMs produced longer responses when producing descriptions of pictures from a city while addressing an audience unfamiliar with that city and used more proper nouns when addressing a familiar audience. In Experiments 2 and 3, similar to previous findings with humans, LLMs used fewer words to describe pictures over the course of a multi-turn interaction. However, this pattern appeared to be sensitive to whether the user prompts also got shorter across turns, suggesting that efficient audience design in LLMs reflects patterns in training data and reinforcement learning, rather than an inherent drive towards least effort.