Our primary purpose in this study was to examine the structure of a response class when new members are acquired through mand training. To do this, we replaced existing mands (e.g., reaching) in three children with autism with two new functionally equivalent mands. Next, we examined their responding under immediate- and delayed-reinforcement conditions. Then, we assessed generalization to novel social partners. We employed a reversal design to examine the effectiveness of mand training and to assess responding under both immediate- and delayed-reinforcement conditions. Our results suggest that all children acquired the new mands and that two of the children emitted these responses as replacements when the social partner did not provide access to the reinforcer contingent on the child’s first mand. Generalization data indicate that all three children emitted the new mands and two of the children alternated between the new mands with novel social partners. We discuss the clinical implications and the conceptual significance of teaching multiple replacement mands to children with autism and severe language delays.