If we have a statement $\sigma$ that we wish to prove in some theory, then usually, the way that we do that is by using the axioms of that theory and the inference rules of our logic system. Once we have done that, we then know that $\sigma$ is true in all models of the theory. The fact that it's true in all models is a consequence of the proof, not a prerequisite for the proof.
You ask: isn't it true that we usually prove things inside a model? The answer is no, absolutely not! In particular, when we prove theorems of ZFC, we almost never have a particular model of ZFC in mind. Theoretically, we are merely applying axioms and inference rules, and there are no models anywhere to be seen.
It is true that when we work in ZFC, we typically imagine that there is some particular collection of objects, $V$, that the axioms are "talking about." However, this $V$ that we have in mind is usually not actually a model of ZFC. It's just an abstract, fictional idea that we use as a mental aid.
In particular, "(class of all sets, $\in$ relation)" is not a model of ZFC. If we already have some model $\mathcal{M}$, then we may use the phrase "the class of all sets of $\mathcal{M}$" to refer to one of the components of $\mathcal{M}$, but the phrase "the class of all sets" all by itself doesn't give us a model.
Admittedly, some mathematicians believe that there actually is one particular model of ZFC that is the "correct" or "intended" model. However, even if we really do have a particular model of ZFC in mind, we nevertheless have to obey a restriction: when we prove things, we're not allowed to use arbitrary properties of the model in our proofs. Instead, we must only use those properties which are axioms of ZFC. By doing this, we guarantee that our proofs hold true in all models of ZFC, not only the particular model we had in mind.