It all boils down to the psychoacoustic model in the encoder. It determines how the bitrate is allocated to various parts of the spectrum, depending on the audio signal.