This study compares antonymous relations between lemmata (such as soft/softer/softest and hard/harder/hardest), words (soft and hard) and word senses (for example, the sense of soft indicating a yielding surface and that of hard describing an unyielding surface). In agglomerative cluster analyses of data from the British National Corpus, specific antonymous adjective senses are found to cluster more tightly and neatly than either antonymous words or lemmata. Moreover, when pairs such as soft and hard co-occur in the corpus, the co-occurring senses are typically instances of antonymous senses that cluster together. This evidence from clustering and co-occurrence suggests that antonymy operates primarily at the level of the word sense, rather than the word or the lemma.