It depends a great deal on what it is you are looking at, and to some extent it is just semantics. But as a rule, if [itex]\lambda[/itex] is your (IR) power counting parameter, then in terms of light cone momenta (and setting the hard scale to 1):
[tex](p^+,p^-,p_\perp)\sim (\lambda,\lambda,\lambda)\Rightarrow{\rm soft}[/tex]
[tex](p^+,p^-,p_\perp)\sim (\lambda^2,\lambda^2,\lambda^2)\Rightarrow{\rm ultrasoft}[/tex]
So that the "offshellness" of a soft mode is [itex]\lambda^2[/itex] while the usoft mode is [itex]\lambda^4[/itex].
When dealing with fully inclusive processes this distinction is irrelevant. The trouble comes in when you have exclusive or semi-inclusive processes. Then you have to distinguish soft from usoft.
See, for instance, Bauer, Pirjol, Stewart, hep-ph/0109045