It's quite expensive, especially if you only need one feature from one package that depends on multiple other GPL packages (which is most of them). Everything included we got quoted over $25K for a feature which was decidedly non-trivial, but the cost was still well above the "let's license it instead of re-implementing ourselves" threshold.
I dare to say that 25K is not that much; if you had to do it yourself, it'd be probably very expensive (of course, I don't know how good you are :-) but you see what I mean).
(and yes, I did 2 years of R&D on quad trees with polygon intersections and "exact computation" is super tough to achieve, you need lots of testing)
25k is actually pretty cheap for this sort of numerical work. That's also licensing about a quarter of the entire library, regardless of how much your actually using.
To put that in context, 25k is roughly the fully loaded cost of a developer month, if they can do this sort of work properly.
Plenty of specialized libraries out there will cost you twice (or more) as much to acquire and a % on your shipped product.
Thinking of this as expensive is either naive, or just the wrong tool for the job.
For the application we were considering and the algorithms we liked from CGAL, exact computation wasn't really an issue. I agree 25K is peanuts if the choice is between developing yourself from scratch or licensing something from CGAL. But in our case the choice was between CGAL, using a simpler and possibly less advanced/less accurate solution, or using an alternative LGPL package. Especially considering the many dependencies most CGAL packages have on other packages, the value proposition for just using a small part of the library for a single algorithm quickly becomes a hard sell...
