Overconfidence is considered to be one of the most prominent behavioral biases, and has been studied with students, executives, and investors using different variants of a confidence interval task with large intervals such as 90%. On average this research has found that individuals are overprecise, leading this type of overconfidence to sometimes be labelled as the “most robust form of overconfidence”. With over 1000 participants, we study different versions of confidence interval tasks, with widths spanning from 10% to 90%. We show that the aggregate finding of overconfidence can be reversed by examining small instead of large intervals. In what we consider to be our most reliable interval task, we find that the magnitude of underconfidence for small intervals is similar to the magnitude of overconfidence for large intervals. Hence, when aggregating across small and large intervals in this task, participants are on average well calibrated. Our results suggest that this form of overconfidence may not be as robust as previously thought.