We consider a continuous-time stochastic model of spiking neurons. In this model, we have a finite or countable number of neurons which are vertices in some graph $G$ where the edges indicate the synaptic connection between them. We focus on metastability, understood as the property for the time of extinction of the network to be asymptotically memory-less, and we prove that this model exhibits two different behaviors depending on the nature of the specific underlying graph of interaction $G$ that is chosen. This model depends on a leakage parameter $γ$, and it was previously proven that when the graph $G$ is the infinite one-dimensional lattice, this model presents a phase transition with respect to $γ$. It was also proven that, when $γ$ is small enough, the renormalized time of extinction (the first time at which all neurons have a null membrane potential) of a finite version of the system converges in law toward an exponential random variable when the number of neurons goes to infinity. The present article is divided into two parts. First we prove that, in the finite one-dimensional lattice, this last result doesn't hold if $γ$ is not small anymore, in fact we prove that for $γ> 1$ the renormalized time of extinction is asymptotically deterministic. Then we prove that conversely, if $G$ is the complete graph, the result of metastability holds for any positive $γ$.