Neurons in a neural circuit have been demonstrated to have astonishing diversity in terms of numbers and targets of their synaptic connectivity and the statistics of their spiking activity. We hypothesize that this is the result of an underlying struggle between reliability, robustness and efficiency of the information represented by their spike trains. Specifically, certain architectures of connectivity foster highly uncorrelated and thus efficient activity, others foster the opposite trends, i.e., robust activity. Both coexists in a neural circuit, leading to the observed long-tailed and highly diverse distributions of connectivity and activity metrics, and allowing the robust subpopulations to promote the reliability of the network as a whole. To test the hypothesis and characterize these architectures, we analyzed several openly available connectomes and found that all of them contained groups of neurons with very different levels of complexity of their connectivity. Using co-registered functional data and simulations of a morphologically detailed network model, we found that low complexity groups were indeed characterized by efficient spiking activity and high complexity groups by reliable but inefficient activity. Moreover, for neurons in cortical input layers, the focus was increasing reliability; for output layers, it was increasing efficiency. To test the effect of the complex subpopulations on the reliability of the network as a whole, we manipulated the connectivity in the model to increase or decrease complexity and confirmed that it affected activity in the expected ways. Our results impact our understanding of the neural code, demonstrating that it is as diverse as neuronal connectivity and activity, and must be understood in the context of the efficiency/reliability struggle.