CST gives you more control over meshing. It's easier to do finer/variable meshing for particular parts of the model than in HFSS. Of course, you have to know what you're doing sometimes - from experience, using a stripline feed not quite aligned with other geometric features can take a while to get meshed correctly, often done manually. HFSS is much easier to mesh with the automatic refinement, but can generate huge meshes.
Narrow vs broadband is huge also. If you have a high-Q structure, don't use CST's time domain solver. If you have a wideband structure, don't use HFSS or CST's frequency domain solver. If you're only interested in a few frequencies, use frequency domain/HFSS.
Both are very accurate and can be just as good in different situations, but sometimes you have to know what you're doing in setting up the simulations. CST can also handle electrically larger structures than HFSS on the same computer, but it'll take a long time to simulate (CST puts a lot of demand on processor, HFSS puts a lot on RAM). A 12G RAM PC was able to simulate one large model in CST (but took days) but unable at all in HFSS.