In a replication and extension of earlier studies by Hermelin & O'Connor, language recoding abilities in autistic, retarded and normal children matched for mental age and digit span, were compared in a verbal recall task. Random word lists, sentences, and anomalous sentences, eight or 12 items in length (for high and low memory span subgroups) were presented and the number of words recalled from each type of input was scored. All low span children recalled sentences better than random lists with normal children superior to retarded and autistic children and the latter group poorer than the retarded group. Autistic children showed a recency effect with both types of input. There were no group differences amongst high span children and sentences were again better recalled than random lists. In Expt II sentences were better recalled than anomalous sentences, with autistic and retarded children equivalent in performance and poorer than normal children. Although low span autistic children were clearly deficient in recall of sentence material when compared with the two control groups, the effect of conditions showed that they were able to use structure to improve recall. Since high span autistic children did not perform differently from controls it is suggested that results from this kind of study may not be generalizable, and that claims for a specific coding deficit in autistic children need further substantiation.