I guess it is matter of opinion and to what degree one can accept the approximation of the linear-response approximation.

I have done two cases studies of it once, I will try to see if I can find them, however I remember a couple of things for sure.

If the system you are calculating on, is typical standard system without spin polarization and perhaps have ohmic resistance,

then it goes something like this.

0-0.2 V Technical no deviation.

0.2V-0.4V Small deviation, but still within the model precision.

0.4V-0.7V Slightly deviation, but still shows the same behavior

0.7V-1.0V The deviation becomes more than 10-20%, and this could be considered high bias.

1.0V - Typical large errors.

However if the target systems is a MTJ's (Magneto-Tunnel-Junction) like FeMgOFe, where you want to understand the details of the spin current

in order to calculate the spin-torque etc, then I would set 0.3 V as a "high voltage".

But there are no need to despair, since one can always use the fully selfconsistent calculate current functionality to get the right results.

Now I will go look, if I can find my case studies on this matter.