Is the provider hiding something?

When the audio contains no speech, does the model still emit text? When you pay for the flagship tier, is it actually different from the cheap tier? When you submit the same clip twice, do you get a suspicious byte-identical response?