GT5's method is definitely where PD are going. GT6 is clearly a stopgap on the visual damage front. With a few more cycles allowed for collision handling, different materials can be better represented and extents and self-intersection (clipping, weird floating bits etc.) better handled at the point of impact. The GT5 E3 demo is probably a good representation of what GT7 will bring.
What's weird is that the GT6 model appears to need more memory. All GT5 needed was the mesh to deform in-place and some procedurally applied dirt / scratches (low-res, so kind of dark dirt). GT6 has apparently higher resolution dirt / damage maps, plus they have a tiling per-vertex normal perturbation map, to give the impression of dents in the reflection mapping and other shaders that utilise the surface normals.
(I'm assuming that car models aren't instanced in GT5, because you can't guarantee that every car won't be a different Premium, so you have to budget your memory for that one worst case. That means each car gets its own copy of vertices to displace, and may also explain why replays "duplicated" damage when rewound etc. The "virgin" geometry must be re-streamed back in after the race and at the re-starting of the replay, or at least only the part that describes the vertices' locations).
Both approaches are innovative, possibly unique in their specific combination of implementations, and reflective of PD's willingness to try bespoke solutions.
I didn't see it mentioned, but I've skipped over the bickering a bit, so I may have missed it, that a major advantage of procedural damage modeling is the fact that separate "damaged" states don't have to be modeled in, nor textured up. That's great in a game with so many cars - its also great for subtle variety and hence an impression of detail. If all cars end up with a high-resolution external mesh in the next game, then we could be in business.
I do wonder how on earth PD are going to get that real-time deformation working with the sub-division ("tessellation"), and even the progressive meshes (the pre-computed splitting and merging of vertices for gradual LoD reduction was done based on their original locations, not their displaced locations...) That's two non-trivial problems, as far as I can tell, and both will probably require yet another unique solution.