Codon usage bias (CUB), where certain codons are used more frequently than expected by chance, is a ubiquitous phenomenon and occurs across the tree of life. The dominant paradigm is that the proportion of preferred codons is set by weak selection. Though experimental changes in codon usage have shown large phenotypic effects, genome-wide population genetics estimates have generally been consistent with the weak selection model. Here we use site frequency spectrum and polymorphism- level data from deep Drosophila melanogaster population genomic sequencing to measure selection on synonymous sites. We find evidence that purifying selection on preferred codons varies in strength from weak to strong (Nes < −10). Our results suggest a new model where the level of CUB in a gene is determined by distribution of selection coefficients across sites. These results also indicate that the functional effect of CUB, and of synonymous sites in general, have been underestimated.