Temporal irregularity and heterogeneities are key features of neuronal spiking activity in cortical networks. Simplified theoretical models of cortical circuits show that these features can be mechanistically accounted for if cortex operates in a fluctuation-driven regime (FDR). In FDR, excitatory and inhibitory currents are approximately balanced, and activity fluctuations emerge intrinsically from the non-linear network dynamics. However, it is still unclear if cortex operates in this regime. We address this question by analyzing sub-threshold membrane potential of neurons in sensory and frontal cortex recorded in decision-making tasks. Sub-threshold activity is highly heterogeneous across neurons in the same neuronal population. For example, the mean distance to spike threshold of the membrane potential varies substantially across neurons and different neuronal populations. The standard FDR framework accounts for the spiking statistics but fails to capture the heterogeneity in the sub-threshold activity, thus challenging the long-standing view that cortex operates in this regime. We extend the FDR framework by introducing a new phenomenological model of point neurons that mimics dendritic integration, with model parameters estimated from simulations of multicompartment neurons. A network consisting of such ‘spatially extended-like’ point neurons can account for both sub and supra-threshold statistics. Our model suggests that neurons in frontal cortex are approximately balanced and operate in the FDR. In contrast, excitatory neurons in Layer 4 of the barrel cortex are mean-driven: they are dominated by inhibition and spike in response to occasional synchronous input (‘mean-driven’). Our work suggests that different populations in cortex can operate in different dynamical regimes. Cortical excitatory neurons closer to the periphery are mean-driven, firing due to strong and correlated external drive, whereas neurons in other populations hover closer to their thresholds, their currents are approximately balanced, and they are driven by input fluctuations.