Ensuring the qualities of modern software systems, such as the Internet of Things, is challenging due to various uncertainties, such as dynamics in availability of resources or changes in the environment. Self-adaptation is an established approach to deal with such uncertainties. Self-adaptation equips a software system with a feedback loop that tracks changes and adapts the system accordingly to ensure its quality goals. Current research in this area has primarily focussed on the benefits that self-adaptation can offer. However, realising adaption can also incur costs. Ignoring these costs may invalidate the expected benefits. We start with demonstrating that the costs for adaptation can be significant. To that end, we apply a state-of-the-art approach for self-adaptation to an Internet of Things (IoT) application. We then present CB@R (Cost-Benefit analysis @ Runtime), a novel model-based approach for runtime decision-making in self-adaptive systems. CB@R is inspired by the Cost-Benefit Analysis Method (CBAM), which is an established approach for analysing costs and benefits of architectural decisions. We evaluate CB@R for a real world deployed IoT application and compare it with the conservative approach applied in practice and a state-of-the-art self-adaptation approach.