Using random class assignment in the China Education Panel Survey, we study the gender achievement gap of lower secondary students in China. Girls outperform boys in Chinese and English by around 0.5 SD, holding all factors including hukou status, cognitive ability and schools attended constant. The Blinder-Oaxaca decomposition reveals that most of the gender gap cannot be explained by observables regardless of the subjects and this is contrary to cognitive scores. Quantile decomposition results indicate that the unexplained “girl premium” is largely monotonically decreasing in the conditional quantiles of the exam score distribution for all subjects, being particularly large at the lower tails, but vanishes altogether at the top decile in Math.