In the particular case of mixing data storage and data manipulation in a single class: It tends to yield classes that continuously accumulate new functionality over time. They also tend to be difficult to refactor int a set of smaller classes. The level of impact on consuming code is very high, because splitting those classes tends to also force major changes to anything that interacts closely with them. And, since all that functionality lived inside a single class, it's likely to be internally resistant to change, too. It's really easy to treat "encapsulation" as a reason to not worry overmuch about shared use of glo^H^H^H instance variables.
Does this kind of stuff mean that OOP didn't turn out to be all it was cracked up to be? Yup. Does this mean we're throwing the whole idea out? Nah, it's just evolving. The original idea's been augmented with other useful ideas like the interface segregation principle, the Liskov substitution principle (mixing behavior with data makes obeying LSP very difficult), and acknowledging the value of referential transparency.
Because it still allows you to write to a higher level of abstraction. So you can have all sorts of different data storage classes - lists, sets, hash maps, etc. - and decide that they all implement an "enumerable" interface, and then any functions that just need to be able to apply some operation to every element in a collection can work with every single one of your collection types.
Haskell has ways of getting us that sort of thing with only structs and bare functions - and does a really nice job of it, too - but most of us aren't working in Haskell.
If you can get over of him calling everything "pure evil", the simplicity he tries to achieve with his controversial view on what is true OOP is quite interesting.
I think that instead of this splitting to data and functions a better solution is to use Single Responsibility, Tell don't Ask, and Demeter's law. This way one can make classes small enough so functionality accumulation won't be a problem and a class will tell what to do only to their "neighbouring" classes. This way every change in the code will be somewhat local. Of course this is an idealistic view which fades with every ORM entity read... :)
Classes can still be parameterized in a way that modules in procedural languages can't, which opens up all sorts of flexibility that wasn't available, or at least was exceedingly awkward to accomplish, in non-OO procedural languages. For example, you're going to want some sort of post-procedural language if you want to use dependency inversion.