The major code improvement is the inclusion of spinors in the VASP code. It is now possible to treat non collinear magnetic structures and spin-orbit coupling on a fully self-consistent basis (see section 6.60)
An automatic way to calculate force constants and vibrational frequencies using finite differences has been implemented (see Sec. 6.21).
For copyright reasons, VASP.4.5 does not support IALGO=8 (M. Teter, Corning and M. Payne hold an US Patent on this algorithm). As a faster and equally reliable substitute for IALGO=8 a Davidson like algorithm has been implemented ( IALGO=38). In addition, it is now possible to select the algorithm using ALGO=Normal, Fast or Very Fast (see Sec. 6.42 for details).
VASP.4.5 also treats the unbalanced lattice vectors differently than VASP.4.4. In VASP.4.4, the charge density at unbalanced lattice vectors was set to zero. But, in combination with US-PPs and PAW potentials, this has significant disadvantages for an accurate description of wavefunctions in the vacuum (STM images). Therefore, the charge at unbalanced lattice vectors is not zeroed in VASP.4.5. To force a behaviour compatible to VASP.4.4, the flag LCOMPAT=.TRUE. can be set in the INCAR file (in VASP.4.4 this flag was used to obtain compatability to VASP.3.2, please do not set this flag if you use VASP.4.4, except if you need compatibility to VASP.3.2).
Additionally, a subtle mistake in the real space projection scheme ( LREAL=.T., LREAL=O,
and LREAL=A) was removed in VASP.4.5.4 and older releases.
The real space projectors are zero beyond a certain radial cutoff 
(line ``Optimized for a Real-space Cutoff X.XX Angstroem'' in the OUTCAR
file). Versions before VASP.4.5.4, however, incorrectly extrapolate the real space projection
operators beyond this cutoff up to /100*101. As a result
the precision of VASP was slightly reduced when using real space
projectors. VASP.4.5.4 and newer releases have removed this error.
Usually the energy differs only by 1 meV per atom, but in some cases
the error can be up to a few meV per atom.
Again compatibility to VASP.4.4 can be forced by
simply setting LCOMPAT=.TRUE. For the few users, who used already VASP.4.5.3,
it is possible to obtain compatibility to that version, by setting only
LREAL_COMPAT=.TRUE. (presently the default is in fact in
any case LREAL_COMPAT=.TRUE.)
Another change concerns the WAVECAR file. To make them smaller, VASP.4.5 writes the WAVECAR file in single precision. VASP.4.5 is still able to read WAVECAR files generated by VASP.4.4, but VASP.4.4 is not able to read files generated by VASP.4.5. If this behaviour is disliked, the pre-compiler flag WAVECAR_double can be specified in the makefiles (Sec. 3.5.14).
Finally, the MPI communication layer and the parallel fast Fourier transformation (FFT) routines have been rewritten to perform optimally on workstation clusters connected by a Fast or Gigabit Ethernet. Usually you can expect a performance improvements of 10-20% with VASP.4.5. Additionally on one processor, the parallel version of VASP.4.5 is now as fast as the serial version.