That makes sense.

Of course, the implementations are going to have varying performance depending on code quality/optimisations but that's a given.

Edit: some other poster gave some cues to my question. I guess you can do some trickery by decomposing the waveform before encoding and recomposing it.

I'll read more on the specs.

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.