I don't think G3D v4 will go away though. If we can just get bump-maps into all the engines, its got a very useful life yet! There is a natural progression in modelling formats and the older formats are more simple and therefore faster to render. So you want to pick the simplest format that supports what you want to achieve for each model.
Most model formats can be pegged to a Doom timeline:
MDL (Quake) and MD2 (Quake2) are equivalent to G3D. The model is one or more meshes, where each mesh has each vertex position per frame and is interpolated to render between frames. I can spot no advantage with MD2 and over G3D v4. Drawing these models is quick and easy - at long range / small on-screen size you can simply calculate the nearest frame and send that to the card without interpolation; closer up you can do your interpolation in a shader.
The problem with these formats is they don't give you forward/inverse kinetics
http://en.wikipedia.org/wiki/Inverse_kinematics.
The solution to kinetics is to have a 'bone' model format. This came with MD5 (Doom3). The model is made from bones / joints and vertices, where each vertex is attached - by various weights - to some of bones. In this way you can have good-looking arms that move without looking like an action man.
So bone formats like MD5 are the high-tech modelling today. They allow very nice movement of humanoids even when those humanoids are drawn very large on the screen. But with an RTS, most units are drawn small and many don't have joint problems.
Rendering a bone format like MD5 is a struggle for shaders and other high-performance paths since the number of weights is per vertex and potentially large. Many engines limit the number of bones that can influence a vertex and end up nop-padding the array of matrices per vertex. All not very satisfactory. So bones in shaders are challenging and going to be slower than a simple interpolation solution. (I would love to be wrong about this.)
MD5 does have additional scene-management animation format which could be very useful for choreographing the interaction of units and buildings especially during build time.
MD3 is a rare format that is half way between MD2 and MD5. Its interesting in that there is export from blender support and I am still trying to understand if it offers the speed of MD2 with the extra turret support that a fuller bone system would give. But its not widely used.
Collada is a bone format (adopted by Unreal Tournament for example) that is an open standard but is notoriously powerful i.e. complex to implement. It is used by the 0ad game.
I think that MD5 would be an excellent
complement to G3D, but that we might have to limit the number of weights supported (pruning on load) and a clever renderer will attempt to redivide the mesh into parts that have n weights and so on. How MD5 might support team colours is still open.
Creating an MD5-equiv bone format for G3D v5 would be perfectly possible and not so scary actually.
