The debate about the impact of routine-biased technical change on wages revolves around the question whether occupational or overall wage distributions polarized. This paper instead argues that routine task prices should decline compared to abstract and manual task prices. I propose a new method, which exploits the sorting of workers into tasks and their associated wage growth in the Roy model, to estimate the changes in task prices under relatively weak assumptions. Empirical results for male workers in two U.S. datasets indicate that task prices polarized during the 1990s and 2000s. The estimates match the upper part of the wage distribution, and they are consistent with differences across countries and time periods in the lower part.