Recent empirical research has shown that improving non-cognitive skills such as grit, conscientiousness and self-control leads to higher academic achievement and better life outcomes. These findings have motivated the creation of educational programs targeting non-cognitive skills around the world. However, both theoretical and empirical works have raised concerns about the reliability of the different methods used to measure non-cognitive skills. The purpose of this study is to compare three methods of measurement of non-cognitive skills. To do so we capitalized on a unique data set collected across 97 middle schools in France, including a student-reported questionnaire, a teacher-reported questionnaire, and a behavioral task. Before analyzing the data, we polled 114 experts in cognitive development and education economics; most experts in both fields predicted that the behavioral task would be the best method. We found instead that the teacher questionnaire was more predictive of students' behavioral outcomes and of their grade progression, while the behavioral task was the least predictive. The three methods of measurement showed similar long-term stability and similar reliability, as measured by Cronbach's alpha. In addition, we show that the behavioral task is the least cost effective method to measure non cognitive skills. Our study suggests that behavioral tasks should not automatically be considered as the best method to measure non-cognitive skills.