It's like writing your own TCP/IP stack instead of using the one included in the OS. Not only is it stupidly redundant, it also means you've lost support, driver updates, ease of use, forward compatibility with new hardware, and regular-user-mode access.
Speed might be a reason to use libusb rather than 30 year old serial I/O libs... Just saying...
On the other hand, libusb is designed for raw USB access, and non-native on at least Windows. But it does add a lot of abstraction which theoretically harms performance. It then goes and does the same things as the "30 year old serial I/O libs" using a non-standard interface. libusb is nice when there are no existing drivers, but totally the wrong tool for these specific devices. Unfortunately, libusb also lacks any support for asynchronous access on Windows too, which makes some device API improvements impractical - before I can move BFGMiner to a completely asynchronous model, I would need to first do some major improvements to the underlying libusb library itself.
Edit: Disclosure... there is one reason I can see using libusb could be beneficial: unfixed bugs in the OS/official OS drivers, or workarounds for buggy hardware. This is the case with Windows's ACM driver (used by ModMiner) - but easily worked around in software (as BFGMiner has done for a while). The chip used in the Icarus also had a bug that prevented it from working with certain USB hosts - this too, was easily worked around in software. But those are the only reasons I can see using libusb would make sense for a device using a serial interface, and they're already managed just fine without it.
Wait, you just said that it was the current, supported, standard interface, and libusb is low level. Then you said libusb adds a lot of abstraction and does the same things as the serial I/O libs, which would make it higher level. Which is it?