This paper examines the cost efficiency of Argentine commercial banks following the financial and monetary reforms of the early 1990s. A panel data set including all commercial banks operating in Argentina during 1996-2000 was used. Cost efficiency is measured using two different techniques. First, input X-inefficiency is estimated via the Distribution-Free Approach, introduced by Berger (1993). Second, financial institutions with different characteristics (private vs. public and domestic vs. foreign) are analyzed based on their cost distribution using quantile regression analysis. In both cases, bank costs are specified as translog functions. When all banks are considered together, empirical results reveal X-inefficiencies representing around 20% of banking costs, which might be interpreted as a reasonable level of inefficiency, as compared to international standards. Analysis of banks sorted by different asset classes indicates that smaller banks tend to depart relatively more from a fully optimizing behavior. On the other hand, according to quantile regression estimations, public banks consistently show (on average) higher costs than private banks, being the differences larger when banks are evaluated at higher quantiles of the (conditional) cost distribution. In other words, the efficient public banks would have similar costs as the efficient private banks, while the inefficient public banks would have higher costs than the inefficient private banks. This could be indicating the unwillingness (or inability) of the government to liquidate public banks operating with high costs. Finally, when banks are distinguished as vs. domestic, cost measures are not significantly altered (on average). Specifically, the estimated quantile function characterizing foreign banks shows costs slightly above domestic banks along most quantile point estimates, and exceptionally lower costs for banks located either at the very lower and upper tails of the cost distribution, although in any case these results (especially the latter ones) turn out to be statistically robust.