Sea ice has a large impact on climatic system and its variability. A good reproducibility of the past state of the sea ice in global climate models will reduce uncertainty in future projection. Here, we present sea-ice simulations for new versions of atmosphere-ocean coupled general circulation models, the Model for Interdisciplinary Research on Climate version 4h (MIROC4h) and version 5 (MIROC5), and assess the reproducibility of the sea ice prior to the future projection. The horizontal resolution of MIROC4h is significantly high for a coupled climate model, although its sea-ice component is based on the previous version. MIROC5 employs some improved schemes including subgrid-scale ice thickness distribution. Hindcast simulations of twentieth-century climate by the new models are compared with observations and with the results of previous versions of MIROC. For the Northern Hemisphere, Arctic sea-ice simulations are improved in both MIROC4h and MIROC5 compared with previous models. MIROC5 generally agrees well with observational data, whereas in MIROC4h, the Arctic sea ice is smaller in summer extent and in thickness. Employment of the ice thickness distribution, which allows large heat exchange through subgrid-scale thin ice regardless of the grid-averaged thickness, and relatively high albedo parameters contribute to reproduce more realistic ice thickness in MIROC5 compared with that in MIROC4h. For the Southern Hemisphere, MIROC4h well reproduces the observed ice edge, especially in winter, while MIROC5 underestimates sea-ice extent. Both models indicate decreasing trends in Arctic sea ice in the late twentieth century. A heat budget analysis of the MIROC5 Arctic Ocean suggests that intensification of ice-albedo feedback accelerates the rate of Arctic ice decline.