Oxygen reduction reaction (ORR) occurs at the cathode of electrochemical devices like fuel cells and in the Huron–Dow process, reducing oxygen to water or hydrogen peroxide. Over the past years, various electrocatalysts with enhanced activity, selectivity, and durability have been developed for ORR. However, an atomic-level understanding of how materials composition affects electrocatalytic performance has not yet been achieved, which prevents us from designing efficient catalysts based on the requirements of practical applications. This is partially because of the polydispersity of traditional catalysts and their unknown structure dynamics in the electrocatalytic reactions. Here we establish a full-spectrum of atomically precise and robust Au_xAg_25−x(MHA)₁₈ (x = 0–25, and MHA = 6-mercaptohexanoic acid) nanoclusters (NCs) and systematically investigate their composition-dependent catalytic performance for ORR at the atomic level. The results show that, with the increasing number of Au atoms in Au_xAg_25−x(MHA)₁₈ NCs, the electron transfer number gradually decreases from 3.9 for Ag₂₅(MHA)₁₈ to 2.1 for Au₂₅(MHA)₁₈, indicating that the dominant oxygen reduction product alters from water to hydrogen peroxide. Density functional theory simulations reveal that the Gibbs free energy of OOH adsorption (∆G_OOH*) on Au₂₅ is closest to the ideal ∆G_OOH* of 4.22 eV to produce H₂O₂, while Ag alloying makes the ∆G_OOH* deviate from the optimal value and leads to the production of water. This study suggests that alloy NCs are promising paradigms for unveiling composition-dependent electrocatalytic performance of metal nanoparticles at the atomic level.