Updating and maintenance of information are 2 conflicting demands on working memory (WM). We examined the time required to update WM (updating latency) as a function of the sequence of updated and not-updated items within a list. Participants held a list of items in WM and updated a variable subset of them in each trial. Four experiments that vary the number of to-be-updated and to-be-maintained items, as well as their positions in the list, are reported. The pattern of latencies was best explained by a model assuming forward scanning of the list, updating modified items, and maintaining nonmodified items. Switching between updating and maintenance incurred a response time cost, which increased with overall set-size. The formation of new item-position associations accounted for an additional response time component. The finding of an update-switch cost provides novel behavioral support for a class of physiologically inspired computational models, in which updating and maintenance require 2 different states of WM.