This paper presents the first benchmark for uncertainty disentanglement, evaluating a wide range of uncertainty estimators across diverse tasks on ImageNet and CIFAR-10. The goal is to disentangle uncertainty into aleatoric (data-inherent) and epistemic (model-based) components. Despite recent theoretical efforts, no existing approach provides disentangled uncertainty estimators in practice. The paper finds that specialized uncertainty tasks are harder than predictive tasks, where performance saturates. It reveals that most existing methods produce highly correlated uncertainty estimates, failing to disentangle aleatoric and epistemic uncertainties. The study highlights the importance of task-specific uncertainty estimators and suggests that combining individual estimators that reflect different types of uncertainty can lead to better disentangled estimates. The results show that methods like Mahalanobis perform well on out-of-distribution detection, while others like EDL and PostNet excel in predictive uncertainty tasks. The paper also demonstrates that uncertainty estimators are robust to distribution shifts, with some methods maintaining performance even under severe corruptions. However, the study emphasizes that no single method is universally effective, and that task-specific approaches are needed for different uncertainty types. The findings suggest that future research should focus on developing more specialized and disentangled uncertainty estimators tailored to specific tasks. The paper provides a comprehensive benchmark of uncertainty methods and tasks, offering insights into the challenges and opportunities in uncertainty quantification.This paper presents the first benchmark for uncertainty disentanglement, evaluating a wide range of uncertainty estimators across diverse tasks on ImageNet and CIFAR-10. The goal is to disentangle uncertainty into aleatoric (data-inherent) and epistemic (model-based) components. Despite recent theoretical efforts, no existing approach provides disentangled uncertainty estimators in practice. The paper finds that specialized uncertainty tasks are harder than predictive tasks, where performance saturates. It reveals that most existing methods produce highly correlated uncertainty estimates, failing to disentangle aleatoric and epistemic uncertainties. The study highlights the importance of task-specific uncertainty estimators and suggests that combining individual estimators that reflect different types of uncertainty can lead to better disentangled estimates. The results show that methods like Mahalanobis perform well on out-of-distribution detection, while others like EDL and PostNet excel in predictive uncertainty tasks. The paper also demonstrates that uncertainty estimators are robust to distribution shifts, with some methods maintaining performance even under severe corruptions. However, the study emphasizes that no single method is universally effective, and that task-specific approaches are needed for different uncertainty types. The findings suggest that future research should focus on developing more specialized and disentangled uncertainty estimators tailored to specific tasks. The paper provides a comprehensive benchmark of uncertainty methods and tasks, offering insights into the challenges and opportunities in uncertainty quantification.