Early on in the campaign, I tried to do a "proper" hexcrawl/sandbox game, in which the players just did whatever the heck they wanted and roamed around the landscape looking for trouble. Although I had some successes with the method, ultimately I felt it didn't work out as well as I'd hoped. This was for several reasons -- I was using an edition of D&D that required significant monkeying to make it fit my vision, and I didn't want to put in the monkeying time, for one -- but the most important, I think, was play time.
My game meets every two weeks, and a typical session probably lasts three hours. Some people start to arrive around six, but by the time everyone's at the table it's half past seven, and we play on a weekday, so people generally don't want to stay out too late. The result is that it's hard to do the kind of marinating-in-the-setting play that involves. I didn't want to weaken the players' freedom of choice by adopting too much of a linear narrative, however.
The result is a sort of ... series of linked mini-sandboxes, I think. There's a developing overall narrative, but I just abstract a lot of the intervening travel time, resource management between sessions, leaving only big NPC relationships and decisions like "are we going to help this guy" or "where shall we go next," and I try to keep each particular situation to no more than one or two sessions. After all, spending four or five sessions in a given place doesn't seem like a lot, but when that translates into two months, it can wear.
I think in some ways I'm still finding the balance between the two zoom-levels. But I'm happy with the way it's going.