Extensible effects are easier to compose in arbitrary ways. If you have a monad transformer stack you can only have each type of transformer once in each stack before you have to start using lift . lift . lift to get to the inner layers as the type class functions will only get you to the outermost instance.

They suffer from performance problems though as Roman discovered at ZuriHac this year: http://ro-che.info/articles/2014-06-14-extensible-effects-fa...