Abstract The purpose of this study was to determine test–retest reliability for peak barbell velocity (Vpeak) during the bench press (BP) and bench press throw (BPT) exercises for loads corresponding to 20–70% of one-repetition maximum (1RM). Thirty physically active collegiate men conducted four evaluations after a preliminary BP 1RM determination (1RM·bw-1 = 1.02 ± 0.16 kg·kg-1). In counterbalanced order, participants performed two sessions of the BP in one week and two sessions of the BPT in another week. Recovery time between sessions within the same week was 48 hours and recovery time between sessions of different weeks was 120 hours. On each day of evaluation the individual load-velocity relationship at each tenth percentile (20–70% of 1RM) in a Smith machine for the BP or BPT was determined. Participants performed three attempts per load, but only the best repetition (highest Vpeak), registered by a linear position transducer, was analysed. The BPT resulted in a significantly lower coefficient of variation (CV) for the whole load–velocity relationship, compared to the BP (2.48% vs. 3.22%; p = 0.040). Test–retest intraclass correlation coefficients (ICCs) ranged from r = 0.94-0.85 for the BPT and r = 0.91-0.71 for the BP (p < 0.001). The reduction in the biological within-subject variation in BPT exercise could be promoted by the braking phase that obligatorily occurs during a BP executed with light or moderate loads. Therefore, we recommend the BPT exercise for a most accurate assessment of upper-body velocity.