In a lexical decision paradigm subjects viewed character strings and indicated if they were true words. Half of the stimulus words occur frequently and half occur infrequently in printed English. The probability that a stimulus was a true word varied across conditions. Event-related brain potentials (ERPs) and reaction time (RT) were recorded. The latency of the P300 ERP was used to estimate stimulus evaluation time independently of factors influenced by response selection and execution. Uncommon words were associated with longer P300 latencies, reduced P300 amplitude, and slower responses. The probability of a stimulus being a true word in any given sequence affected RT, but produced only small effects on P300 latency. The results suggest that: (1) the effects of word frequency originate early during processing, (2) stimulus probability affects response priming mostly for stimuli which do not evoke a high-level code, and (3) uncommon words require more processing capacity for their evaluation than common words. The results are discussed in the context of word recognition models.