H uman studies show that uninvolved bystanders often pay to punish defectors (1, 2). Such behavior has typically been interpreted in terms of group-level benefits (3) despite theoretical predictions that third-party punishment can yield individual benefits to the punisher (4). Here, we show that male cleaner fish, Labroides dimidiatus, punish their female partners if females cheat while inspecting model clients. Punishment promotes female cooperation, yielding direct foraging benefits to the male. This finding demonstrates empirically that third-party punishment can evolve via self-serving tendencies.When servicing clients, cleaners either cooperate by removing ectoparasites or cheat by taking a bite of mucus. Cleaners prefer to eat mucus (5) but risk clients terminating the interaction if they do (6). The conflict of interest between cleaner and client is less pronounced during inspections by cleaner pairs than during singleton inspections (7). Increased service quality during pair inspections was interpreted as resulting from female cooperation, both under natural conditions and in the laboratory where Plexiglas (acrylic plastic) plates serve as model clients. Larger males regularly chased smaller female partners if females caused clients to leave, particularly in the laboratory (7). This male behavior is peculiar because clients, not males, are the victims of biting females. Despite this fact, it is possible that male chasing could cause females to behave more cooperatively in future interactions. If males benefit personally through increased foraging success, then the males' behavior represents third-party punishment that evolves via self-regarding, rather than altruistic, tendencies.Cleaners were trained to feed off model clients (Plexiglas plates) containing two food types: preferred prawn and less-preferred fish flake (8). Eating prawn corresponded to cheating the client by eating mucus, whereas eating flake corresponded to cooperating by removing ectoparasites. Eating prawn resulted in the immediate removal of the plate. We replicated earlier results (7) that males more often chased females when the latter's behavior caused plate removal (Wilcoxon signedrank test, n = 8, V (rank sum of positive values of x = y) = 21, P = 0.036). After being chased, females were less likely to eat prawn again in the second plate presentation 60 s later (Wilcoxon signed-rank test, n = 8, V = 34, P = 0.030; Fig. 1A) and fed against their preference, eating a higher ratio of flake to prawns (Wilcoxon signedrank test, n = 8, V = 36, P = 0.008; Fig. 1B). This allowed males to increase their food intake in the second presentation (Wilcoxon signed-rank test, n = 8, V = 28, P = 0.02; Fig. 1C). Conversely, when chasing was prevented with a temporary partition, females that cheated in the first presentation were equally likely to cheat in the second one (Wilcoxon signed-rank test, n = 8, V = 11, P = 0.67; Fig. 1A) and did not eat a higher ratio of flake to prawn items (Wilcoxon signed-rank test, n = 8, V = 8, P = 0.35; Fig. 1B), a...