The phenomenon that synonymous codons are used with different frequencies (also known as codon usage bias) exists in a wide range of organisms. It has been extensively reported that codon usage bias is positively correlated with gene expression level, which has been explained by the natural selection on translational accuracy and/or efficiency to promote codon usage bias in highly expressed genes, although they are both challenged by recent studies. In this study, we investigated an opposite possibility that codon usage bias can directly regulate the mRNA level. We generated two synthetic GFP libraries of in total 3,556 variants that varied at 12 synonymous sites, and measured the mRNA levels of these variants. Unexpectedly, we observed that genes using a higher proportion of preferred codons tend to exhibit higher expression levels, suggesting that codon usage bias plays an important role in regulating gene expression. In other words, the correlation between codon usage bias and expression level could exist even in the absence of natural selection. Our study demonstrates the pleiotropic function of codon usage bias, provides a new explanation for the relationship between codon usage bias and gene expression level in the genome, and paves the road for synthesizing of artificial organisms.