Theoretically, the number of samples per pixel will determine your bandwidth increase -- Basically, 2xMSAA should require twice the amount of BW as 0xMSAA.
However, modern gfx cards generally compress the frame buffer data when running in MSAA, so that the cost of BW is usually much lower; 2xMSAA is often close to free, while 4xMSAA is usually a reasonably low cost. But it's highly dependant on the compression algorithm used, as well as the frame being rendered. For example, a full screen quad will generally cost no more in MSAA as in non-AA cases.
As for fillrate, the typical Z hit is (1/number of samples). ATI does do better than this on current HW (2x that), but it's a good general rule, as you do have to do Z compares for all fragments. Color blending can be similar, dependant on the color compression algorithm used.
Finally, in terms of quality, it's generally subjective, but it's fair to say that 2xMSAA or 4xMSAA will look better for edges (at 10x7) than 12x10 no AA, since there are effectively many more samples involved. However, texture content will look better at 1280 than 1024, so the overall image might or might not look better. It can go either way, depending on the image displayed. Better to run at the highest resolution your monitor does well, with as much AA as you can afford, fps wise.