Happy, surprised, disgusted, angry, sad, fearful, and neutral faces were presented extrafoveally, with fixations on faces allowed or not. The faces were preceded by a cue word that designated the face to be saccaded in a two-alternative forced-choice discrimination task (2AFC; Experiments 1 and 2), or were followed by a probe word for recognition (Experiment 3). Eye tracking was used to decompose the recognition process into stages. Relative to the other expressions, happy faces (1) were identified faster (as early as 160 msec from stimulus onset) in extrafoveal vision, as revealed by shorter saccade latencies in the 2AFC task; (2) required less encoding effort, as indexed by shorter first fixations and dwell times; and (3) required less decision-making effort, as indicated by fewer refixations on the face after the recognition probe was presented. This reveals a happy-face identification advantage both prior to and during overt attentional processing. The results are discussed in relation to prior neurophysiological findings on latencies in facial expression recognition.