We study production networks of automotive assembly plants to shed new light on the impact of flexibility on plant productivity. We observe that the contemporary manufacturing networks of automotive assembly plants of three US companies (Chrysler, Ford and General Motors) have become less flexible over the years despite the reported benefits of flexibility. To understand this phenomenon, we utilize flexibility indices that have been developed in the modeling literature to measure flexibility but have never been tested in an empirical setting. We identify shortcomings in existing indices and propose a new index to measure network flexibility. Using our proposed index, we find that both extremes of flexibility (too much or too little) affect productivity negatively. Therefore, intermediate levels of flexibility are optimal because they balance the trade-off between better matching of supply and demand with excessive downtime due to model changeovers (changing production from one model to another), which were not accounted for in the modeling literature. Using plant-level production schedules, we find that productivity losses due to changeovers have a significant negative effect on manufacturing productivity, often making celebrated “chaining” approaches to network configuration uneconomical. Counterintuitively, firms can often benefit by reducing flexibility levels, depending on the current level of flexibility in the manufacturing network. For example, our estimates indicate that a firm with a highly flexible production network can gain up to 8.8% in productivity by rearranging its network. This can result in an average savings of 460,000 labor-hours in a plant with an average production of 200,000 vehicles per year. Using simulation, we show that well-established long-chain configurations cease to perform better than sparser configurations when changeover losses are accounted for, indicating that firms can be better-off by adopting a sparser structure than chaining to improve productivity.