Based on my admittedly limited research, it seems like you should use Laya for much more deterministic tasks where you have some training data. It won't be as good as Jev for zero shot cases.

I think it can make a lot of sense to make a set with jev and then train laya on that set and do the rest with laya for saving $

agree, too good to be true for one shot cases