Theoretical and experimental studies of assessment in animal contests have, until now, focused on disputes between single individuals. However, whereas single competitors usually avoid fights with opponents that are larger or stronger than themselves, in contests between social groups competitors might be expected to adjust their agonistic behaviour according to the number of individuals in their own and the opposing group. This hypothesis was tested using playback experiments to generate controlled artificial contests between groups of female lions. Recordings of single females roaring and groups of three females roaring in chorus were played back to simulate the presence of unfamiliar intruders within the territories of 21 different lion prides in Serengeti National Park, Tanzania. Defending adult females were less likely to approach playbacks of three intruders than of a single intruder and on occasions when they did approach three intruders they made their approach more cautiously. Defenders also carefully adjusted their decision to approach according to the size and composition of their own group, and attempted to recruit extra companions to the contest by roaring when some were absent at the time of playback. A strong selective advantage to avoiding the costs of fighting with larger groups could have led to the widespread evolution of numerical assessment skills in social species.