As BlasterN suggested you should use mipmapping - increase the size of the cube texture so that the top level is larger (for when you're close to it), then when you're further away the lower mip levels will be used.
You could define 2 cube maps and use material LOD to switch between at distance X if you liked, but a single cube map of the larger size would actually take less space (assuming all textures has mipmaps enabled anyway). Plus with mipmapping the transition is smoother.
Also, a bit `off topic`: is it possible [yet] to dynamically project what's around the said reflected texture onto it as a reflection / cubemap
Yes, you use 6 render to textures, each to a face of the cube map surface. Just create the cube map with the TU_RENDERTEXTURE flag and attach cameras to the render targets for each face. wumpus posted a working example of it a few months back iirc.