I don't think you can say that SPH is always much less costly than continuum methods. Less costly than implicit DNS methods, perhaps, but explicit ones can be faster. These have their own problems though, mainly that it can be tricky to predict how the time step will be chosen according to the CFL criterion. That makes them at least pretty useless for real-time applications – you can't rely on it that the needed frame rate will be fulfilled. SPH in principle also needs well-chosen time steps, but like with implicit DSL these aren't all-important in the sense that the solver goes completely berserk when you exceed it – you'll merely get a solution that doesn't really obey the physics equations.
This is probably the main reason that the different approaches are differently popular in science and graphics: in science and engineering, you want results that can well be reasoned about. You want to be able to evaluate local quantities in absolute terms, confirm some theoretical prediction within given error bounds, etc.. DNS (even though it's often enough still rather shaky theoretically speaking, “we've tried two different resolutions and the result agrees, probably this is close enough to the exact solution”) can fulfill those goals pretty well.
In graphics, you scarcely care about such quantitative concerns. What you care about is that it should look good and in particular without obvious artifacts. Both DNS and SPH do cause artifacts, but in DNS they tend to be visually “obviously digital” blocking etc. effects. For SPH it'll be more of a randomly-scattered blurring. For science that's actually a bit of a problem because it can be tricky to predict and recognise, but for graphics that's an advantage because you really don't want the viewers to notice it when the results are unphysical, unlike in science where physical correctness is the most important thing and it's actually good if artifacts are visually obvious.