We analyze the cross-national distribution of GDP per capita and its evolution from 1970 to 2003. We argue that peaks are not a suitable measure for distinct growth regimes, because the number of peaks is not invariant under strictly monotonic transformations of the data (e.g. original vs. log sc… ale). Instead, we model the distribution as a finite mixture, and determine its number of components (and hence of distinct growth regimes) from the data by rigorous statistical testing. We find that the distribution appears to have only two components in 1970-1975, but consists of three components from 1976 onwards. The level of GDP per capita stagnated in the poorest component, and the richest component grew much faster than the medium component. These findings empirically confirm the predictions of the unified growth theory.