Dear Eduardo, thanks for your attempts to clarify the situation!
Aparently the universes are not closed in the sense that:
1) X ~ Y, Y belongs U ===> X belongs U
which is currently accepted (as we see in Thomas posting above) in the naive practice of category theory with universes.
I certainly did not incline this (despite my partial involvement into HoTT :-)) I rather tacitly assumed that a locally U-small means that all homsets are elements of U. This I find natural though I read that Grothendieck and Verdier formulated it much more liberally. However, using AC every locally U-small category is isomorphic to some category where all hom-sets are elements of U. Moreover, every category equivalent to such a category is locally U-small in the liberal sense of Grothendieck and Verdier. Now, for a category C which is locally U-small in the restricted sense [C^op,U] is locally U-small in the restricted sense as shown by Paul's argument. But if C' is locally U-small in the liberal sense then [C'^op,U] \cong [C^op,U] and the latter is locally U-small in the restricted sense and thus [C'^op,U] is locally U-small in the liberal sense. Thus, I think in SGA4 they made a mistake. No question that both guys were great mathematicians but that doesn't prevent them from making little mistakes in logic. Thomas [For admin and other information see: http://www.mta.ca/~cat-dist/ ]