Generally speaking neural networks and genetic algorithms are not used in games, and probably not that often outside of games either.
The main reason these are taught in AI academia is not because of their practical applicability but because they are quite easy to explain as teaching devices - both have mathematical and biological analogues that allow a student to understand how they could work.
In the real world, you typically need reliability and predictability. The problem with learning methods is that if they learn 'in the wild' then they can learn the wrong patterns and be unreliable. An NN or a GA could potentially reach a local maximum which is not guaranteed to be good enough to provide the gameplay experience required, for example. Other times, it might end up being too good, finding a perfect strategy that is unbeatable. Neither is desirable in most entertainment products.
Even if you train offline (ie. before launch, and not during gameplay), an apparently good looking set of data could be hiding anomalies that, once found by a player, are easy to exploit. A neural network in particular typically evolves a set of weights which is quite opaque to study, and the decisions made by it are hard to reason about. It would be difficult for a designer to tweak such an AI routine to perform as desired.
But perhaps the most damning problem is that GAs and NNs are generally not the best tools for any task. While good teaching devices, anybody with sufficient knowledge of the subject domain is generally better equipped to use a different method to achieve similar results. This could be anything from other AI techniques such as support vector machines or behaviour trees through to simpler approaches such as state machines or even a long chain of if-then conditionals. These approaches tend to make better use of the developer's domain knowledge and are more reliable and predictable than the learning methods.
I have heard however that some developers have used neural networks during development to train a driver to find a good route around a racetrack, and then this route can then be shipped as part of the game. Note that the final game does not require any neural network code for this to function.
The 'cost' of the method is not really the problem, incidentally. Both NNs and GAs can be implemented extremely cheaply, with the NN in particular lending itself to pre-calculation and optimisation. The issue is really that of being able to get something useful out of them.