Codon usage bias is a ubiquitous phenomenon, which may be caused by mutational bias, selection, or both. The patterns of codon usage in plants are not well understood. Datasets of expressed sequence tags (ESTs) available for many plant species provide the resources for large-scale comparative analysis of codon usage patterns. We developed a computational approach to translate EST or assembled contig sequences, and then used the coding information for comparative analysis of codon usage in 12 plant species, including 6 eudicots, 5 monocots and the green alga Chlamydomonas reinhardtii. While codon nucleotide composition is highly conserved within eudicots or monocots, there is a significant difference between these two major taxonomic groups of higher plants. The third nucleotide position of codons is AU-rich in the eudicot genomes (35-42% of G+C content), but GC-rich in the monocot genomes (59-61% of G+C content). To identify optimal codons in these species, we used EST counts to estimate gene transcript levels. It was demonstrated that codon usage bias is correlated positively with gene transcript levels. Interestingly, the use of optimal codons appears to be well conserved between eudicots and monocots, and to a lesser degree between the higher plants and C. reinhardtii. Most of the optimal codons end with a C or G base, regardless of the different nucleotide composition in these genomes. The results suggest that plant codon usage is affected by translational selection, and the selective pressure appears to be conserved in the plant kingdom.