I have concluded that starting a (reasonably high poly) model with a subdivided cube is always better than anything else. Even if I have to model a pipe, or a cone a cube is the better choice.
In the case of the pipe, if I start with a cylinder I have to commit to the resolution from the beginning, since Catmull-Clark subdivision doesn't work well on it and simple doesn't preserve curvature. It's also harder to create bends than with a subdivided cube where a single 45° loop makes for a nice 90° bend.
In case of the cone, the inability to create edge loops is enough to make the subdivided cube a better proposition.
The spheres have one or both of the above disadvantages depending on the type and orientation.
I have little experience with torus-like objects so I can't say anything about them.
Is the above conclusion right? If not, please explain why, or in which cases it is better to start with which primitive.