When should I use gdal_merge.py?

Question

When building a single georeferenced image from a collection of tiles, gdalbuildvrt followed by gdal_translate is much faster than gdal_merge.py.

In my own measurements of

gdal_merge.py \
        --config GDAL_CACHEMAX 3000 \
        --optfile filelist.txt \
        -o $output_file

and

gdalbuildvrt \
        -input_file_list filelist.txt \
        $tmpdirname/out.vrt
gdal_translate \
        $tmpdirname/out.vrt \
        -of GTiff \
        $output_file

on a collection of 46 GeoTIFFs of 600×600px each, I got over 30min vs. 1min 58s.

It seems strange that the detour via a virtual dataset is so much more performant. Does it have any downsides?

When should I use gdal_merge.py instead, anyway? (And if there aren't any cases for that with gdal_merge.py's current implementation, why isn't gdal_merge.py instead just implemented as an alias to these two actions?)