With a quality normal map and a good texture, you can achieve a visually un-noticable difference 99% of the time if you know what you're doing. I can't imagine a time where you would want to purposefully limit yourself to the amount of enemies or map detail you would have on-screen, nor could I imagine a time where you would want to exchange polygons for these items in place of a few extra bumps and nobs on a weapon that the user won't pay attention to.
I understand the concept of normal maps, I know they only adjust the texture lighting to give an effect of detail, but this is all that is needed in most cases. The way I see it, if a part of your mesh is going to bump out more than it's width or height, then it should be modelled, otherwise a normal map will suffice.
Comparing polygon counts to models from next gen games like COD is unreasonable as well. If this is for use in FPSC, or a DBP-made game, you should always be aiming lower. The source engine and UT3 engines have been rediculously optimized in the render pipeline to speed things up, and being programmed in a lower-level language grants them even more speed.
The reason these weapons of 5-10k polygons do work in most FPSC games - from what I've seen - is because all of the map objects are very low-poly. This is fine, I completely agree that if your weapon is going to take up a third of your screen space then more detail is necessary, however it's redundant to go over-the-top and pack the weapon full of polygons when these could be much better used for adding more environmental detail while still maintaining a weapon model that looks unrecognizably different to the common gamer.
Take a look at this image:
Aside from the barrel, is there anything noticably different between the 5218 tri and 3776 try versions? This is what I'm talking about; polygons have been added just for the sake of adding polygons. I bet with the right normal map and texture, the different between the 5218 version and the 1804 version would still be unnoticable aside from increasing the polycount on the barrel, which would only bring the entire count up by 100 tris max.
Another point about the image; in the 3k version there are prime examples of what a normal map could simulate perfectly. Those ridges on the pump / hand grip, the circle notch above the trigger, and the inset grooves running the length of the back of the weapon, all could be simulated with a normal map bearing no visually noticable change.
