Large-scale cDNA sequencing projects and tiling array studies have revealed the presence of many unannotated genes. For protein coding genes, small coding sequences may not be identified by gene finders because of the conservative nature of prediction algorithms. In this study, we identified small open reading frames (sORFs) with high coding potential by a simple gene finding method (Coding Index, CI) based on the nucleotide composition bias found in most coding sequences. Applying this method to 18 Arabidopsis thaliana and 84 yeast sORF genes with evidence of expression at the protein level gives 100% accurate prediction. In the A. thaliana genome, we identified 7159 sORFs that are likely coding sequences (coding sORFs) with the CI measure at the 1% false-positive rate. To determine if these coding sORFs are parts of functional genes, we evaluated each coding sORF for evidence of transcription or evolutionary conservation. At the 5% false-positive rate, we found that 2996 coding sORFs are likely expressed in at least one experimental condition of the A. thaliana tiling array data. In addition, the evolutionary conservation of each A. thaliana sORF was examined within A. thaliana or between A. thaliana and five plants with complete or partial genome sequences. In 3997 coding sORFs with readily identifiable homologous sequences, 2376 are subject to purifying selection at the 1% false-positive rate. After eliminating coding sORFs with similarity to known transposable elements and those that are likely missing exons of known genes, the remaining 3241 coding sORFs with either evidence of transcription or purifying selection likely belong to novel coding genes in the A. thaliana genome.