Utilizing a crustal velocity model that includes the complexity of the irregular subsurface structure of the Kanto basin, we have performed three-dimensional (3D) finite-difference (FD) simulations of near-source long-period strong ground motions in the Tokyo metropolitan area for the 1990 Odawara (MJ 5.1) and the 1923 Kanto (MS 8.2) earthquakes. Constraints on the development of the 3D velocity model come from available geological and geophysical data, as well as our previous 1D waveform modeling results (Sato et al., 1998a). The simulation of the moderate-sized Odawara earthquake demonstrates that the 3D velocity model works quite well at reproducing the recorded long-period (T > 3.33 sec) strong motions, including basin-generated surface waves, for a number of sites located throughout the Kanto basin region. Using this validated 3D model along with the variable-slip rupture model of Wald and Somerville (1995), we then simulate the long-period (T > 4 sec) ground motions in this region for the 1923 Kanto earthquake. The simulation results for the 1923 event show that the largest ground motions occur east of the epicenter along the central and southern part of the Boso Peninsula. These large motions arise from strong rupture directivity effects and are comprised of relatively simple, source-controlled pulses with a dominant period of about 10 sec. North of the epicentral region, in the Tokyo area, 3D basin-generated phases are quite significant, and these phases produce large-amplitude late-arriving pulses in the ground motions. At station Hongo (HNG), which is the only site having digitized and restored near-fault strong-motion records for this event, our 3D simulations compare quite well with the ground motions of the restored Imamura seismogram. For the restored Ewing record, our 3D simulations reproduce the phase and amplitude of the initial pulses of motion; however, the dominant period of the large-amplitude later phases is noticeably shorter in the simulations (about 5 to 6 sec) than in the observation (13 sec). These results suggest that the restored Imamura seismogram may be a better representation of the gross features of the actual motion than the restored Ewing seismogram, although the first (clipped and restored) part of the Imamura seismogram may still underestimate the strength of the actual motion.