Methods and apparatus to optimize package level power state usage are described. In one embodiment, a processor control logic receives a request to enter a lower power consumption state (such as a package level deeper sleep state). The control logic determines the time difference or delta between a last entry into the lower power consumption state and the current time. The control logic then causes the flushing of a last level cache based on a comparison of the time difference and a threshold value corresponding to the lower power consumption state. Other embodiments are also claimed and disclosed.