In this paper, we present an improved method of anchor models for speaker verification. Anchor model is the method that represent a speaker by his relativity of a set of other speakers, called anchor speakers. It was firstly introduced for speaker indexing in large audio database. We suggest a rank based metric for the measurement of speaker character vectors in anchor model. Different from conventional metric methods which consider each anchor speaker equally and compare the log likelihood scores directly, in our method the relative order of anchor speakers is exploited to characterize target speaker. We have taken experiments on the YOHO database. The results show that EER of our method is 13.29% lower than that of conventional metric. Also, our method is more robust against the mismatching between test set and anchor set.