Some durations such as those of block trades may have the properties of both heavy tails and extreme values. To model such type of data, we suggest the Fréchet distribution for the innovations of the autoregressive conditional duration (ACD) model, and hence the Fréchet ACD model. Some statistical inference tools including the maximum likelihood estimation and diagnostic tools for model adequacy are derived, and their finite-sample performance is evaluated by Monte Carlo simulation experiments. The usefulness of the new model is demonstrated by analyzing the durations of block trades on two stock exchanges.