Model-based reinforcement learning (MBRL) holds the promise of sample-efficient learning by utilizing a world model, which models how the environment works and typically encompasses components for two tasks: observation modeling and reward modeling. In this paper, through a dedicated empirical investigation, we gain a deeper understanding of the role each task plays in world models and uncover the overlooked potential of more efficient MBRL by harmonizing the interference between observation and reward modeling. Our key insight is that while prevalent approaches of explicit MBRL attempt to restore abundant details of the environment through observation models, it is difficult due to the environment's complexity and limited model capacity. On the other hand, reward models, while dominating in implicit MBRL and adept at learning task-centric dynamics, are inadequate for sample-efficient learning without richer learning signals. Capitalizing on these insights and discoveries, we propose a simple yet effective method, Harmony World Models (HarmonyWM), that introduces a lightweight harmonizer to maintain a dynamic equilibrium between the two tasks in world model learning. Our experiments on three visual control domains show that the base MBRL method equipped with HarmonyWM gains 10%-55% absolute performance boosts.