Although the prediction performance is crucial for a classifier, its cost of use is also an essential issue for practical application. The aim of this article is to propose a prediction method that controls not only the error rate but also the cost of the construction of the classifier. The main idea is that some examples are easier to predict than others and can be predicted using fewer variables i.e. with a lower prediction cost. Our method, called CASCARO, is based on a cascade of reject classifiers of increasing cost. The first classifier of the cascade required only one variable, if the prediction is not reliable the second classifier requiring one more variable is used. The principle is repeated until the last classifier using all variables. This type of cascade raises two scientific problems: the structure of the cascade (the order of the classifiers) and the simultaneous computation of the rejection regions of the classifiers. The experiments show that CASCARO produces significant improvements in the use cost without decreasing prediction performance.