The development of in vitro neural networks depends to a large extent on the scaffold properties, including the scaffold stiffness, porosity, and dimensionality. Herein, we developed a method to generate interconnected neural clusters in a multiscale scaffold consisting of a honeycomb microframe covered on both sides with a monolayer of crosslinked gelatin nanofibers. Cortical neural precursor cells (NPCs) were firstly produced from human induced pluripotent stem cells (hiPSCs) and then loaded into the scaffold for a long period of differentiation toward cortical neural cells. As result, neurons and astrocytes self-organized in the scaffold to form clusters in each of the honeycomb compartments with remarkable inter-cluster connections. These cells highly expressed neuron-and astrocyte-specific proteins, including NF200, Tau, Synapsin I and GFAP, and showed spatially correlated neural activities. Two types of neural clusters, i.e., spheroid-like and hourglass-like clusters, were found, indicating the complexity of neural-scaffold interaction and the variability of three-dimensional neural organization. Furthermore, we incorporated a reconstituted basement membrane into the scaffold and performed coculture of the neural network with brain microvascular endothelial cells. As a proof of concept, an improved neurovascular unit model was tested showing large astrocytic endfeet on the backside of the endothelium.