I'd say they have reliability issues but for the price it's worth it.

I like that usage isn't measured per token but per computation time, which means that you get more usage when models become more efficient.