ECLiPSE and opm-autodiff seem to be in that range of accuracy, too. At least the number of Newton iterations per time step now matches that of autodiff quite well...