I personally like how Ashikhmin-Shirley looks since the diffuse term changes with the view angle. This is true for Oren-Nayar as well but it is less noticeable.
We presented a more accurate and more efficient Oren-Nayar last year at SIGGRAPH (see links provided by Matias Goldberg: “Beyond a Simple Physically Based Blinn-Phong Model in Real-Time (+ slides)”)
I am not so much into Cook-Torrance as it seems to result in overpowering speculars, at least for me. Since I don’t have a way to actually specify reflectance at normal incidence on my models for my engine yet it could be that I just don’t have a very good conversion of specular power to reflectance at normal incidence. But checking images online they do also seem to have much wider and area-encompassing speculars.
I would use our improved Oren-Nayar model by default for low-end devices.
If computing power is removed from the equation I would stick with Ashikhmin-Shirley coupled with the layered material system that was also presented in the same link as above (of course, only when appropriate—not all surfaces need a layered-material system).
Here are results of Ashikhmin-Shirley:
This shot in particular sold me on Ashikhmin-Shirley with the nice transition of oranges over the hood of the car.
No ambient-occlusion was used on the bumper. It is just the natural result of Ashikhmin-Shirley.