Well, that depends on the application. For astronomy and politics within a couple of orders of magnitude is often enough. Engineering and statistics (except where quoted by politicians) like to be within an order of magnitude on the "safe" side. In maths you can often get a little closer than that.