Normally you do the reflection/half vector calculation in the vertexshader anyway, so performance shouldn't be a problem - no matter what model you use.

Take a look at both models and take the one that looks better to you.