These narrow evaluations create the appearance that the open-source models outperform proprietary ones