The way you can think of it is that in OCaml everything is implicitly wrapped in an IO monad. In Haskell the IO monad is explicit, so if a function returns something in IO you know it can perform input and output, in OCaml there is no way to tell just from the types. That means that in Haskell the code naturally stratifies into a part that does input and output and a pure core. In OCaml you can do the same thing, however it needs to be a conscious design decision.

True Functional Programmers will probably correct my lack of nuance here, but my understanding is that the IO monad is basically just scribbling some category theory formalities on top of IO ops so that everything is still technically a pure function. You can think of the IO monad as representing the "state" of the rest of the universe outside of your program, which your program reads or modifies. As you pass through your monadic bind ops (i.e., as you read or write IO), the state is carried through implicitly and "modified" as appropriate. All functions, then, are just transforming data (i.e., either your program's data or the rest of the universe), so everything is pure.