Recent experiments have revealed that visual cortex operates as an inhibition-stabilized network (ISN), in which unstable excitatory coupling is balanced by strong inhibition. Computational models like the stabilized supralinear network (SSN) describe the mechanisms behind this regime, and also postulate that a transition occurs from non-ISN to ISN dynamics with increasing visual input. However, while some studies find input-dependent effects that support a shift in dynamical regime, others have shown ISN-like dynamics in the spontaneous state, leaving open questions about the existence and nature of such a transition, and how it might be measured experimentally. To resolve these unknowns, we combined photostimulation experiments in mouse primary visual cortex and computational network modelling to study cortical dynamics over different visual input levels. We performed simultaneous all-optical recordings and stimulation of groups of L2/3 excitatory neurons to infer the influence of each stimulated neuron on the local excitatory network. Focusing on this influence as a function of signal correlation, we observed a shift from an excitatory influence at spontaneous and low-input regimes to an inhibitory influence with high visual input. To elucidate the circuit mechanisms behind these results, we modelled the perturbation experiments using a rate network. We observed that this shift in influence can be explained by an increase in synaptic connectivity efficacy as a function of visual input. Finally, we found that an SSN-like power-law nonlinearity could act as a plausible mechanism behind this efficacy change. Together, these results show that the dynamical regime of mouse visual cortex becomes more inhibition dominated as visual input increases, transitioning from excitatory cooperation towards more inhibitory competition. Moreover, this dynamical regime change is feature-specific, which could explain why it is not observed in population-level perturbation experiments, thereby reconciling previously conflicting results and supporting the SSN as a plausible model of cortical dynamics.