Seasonal predictability of global surface air temperature for the 100 years of 20th century is examined using the Climate of the 20th Century international project (C20C) AGCM experiment. The C20C experiments reproduce reasonably well the observed warming trend over the globe. The perfect model concept, one simulation being considered as observation, is utilized to examine the changes of seasonal mean predictability for the last 100 years. The global pattern correlations of seasonal mean temperature show clearly the seasonal mean predictability being increased since 1920s. The analysis of the ensemble mean and deviation also shows that the signal to noise ratio is much increased for the recent 30 years, particularly in the tropical and subtropical Pacific. The increase of the seasonal predictability is found to be related to the enhancement of SST variability over the tropical Pacific, which appears to be related to the global warming.