Concerning the recognition of closely-connected Completely Automated Public Turing Test to Tell Computers and Humans Apart (CAPTCHA), a recognition algorithm based on spectral-clustering Recurrent Neural Network (RNN) ensemble was proposed. This algorithm firstly used disagreement measure for distance between two RNNs, thus constructed a graph composed by candidate RNNs. Then, a graph cluster method was used to divide RNNs into clusters. Finally, the best RNN in each cluster was selected. The experimental results reveal that: compared with single candidate RNN, recognition rates of this algorithm is increased by 16%. Compared with the ensemble of all candidate RNNs, ensemble size of this algorithm is much smaller, it is about 23% of the original size.