I don't really see most of the changes as improvements.
Having a different, explicitly-noticeable syntax for mutable updates is nice, because it calls out mutability (which should be used sparingly).
I don't see extra braces as necessarily an improvement, given that OCaml's local scopes are already quite unambiguous thanks to "let ... in". On that note, Removing "in" and just going with semicolons removes another "smelly-code-callout" by making it less obvious what's imperative and what's functional.
I actually don't like ambiguity between type annotation and value assignment in my records. It's clear in current OCaml that {a: int} is a type declaration and {a = 1} is a value declaration/assignment. Moving to colons-for-record-values is at best a bikesheddy, backwards-incompatible change for change's sake, and at worst a breaking-change way of code less clear.
Speaking of making code less clear, how is "int list list" not clear? It's an int-list list. As in, a list of int-lists. So of course it should parse as "(int list) list". Why change to backwards annotations? Just to prevent existing code from working as-is, and making people used to reading ML spend extra brain cycles on remembering that your types read the opposite way?
And they make a huge deal out of their type for tuples being "(a, b)" instead of "(a * b)". Yeah, okay, I get it. It's not that big a deal, since people are used to reading product types as, well, products.
The other thing that seems weird to me is the need to change to a "fat arrow" instead of a "skinny arrow", again for no real reason. In fact, it just makes it more likely that you'll confuse it with a comparison operator. Nobody tries to type ">-", but people try to type ">=" all the time. You're just switching for the sake of switching, and it's not an improvement.
Their example code of their replacement for match...with is especially egregious. If you showed me the OCaml snippet and the Reason snippet unlabelled, I would think that the OCaml snippet is the new-and-improved version, since it's much more compact, much less noisy, and reads more like what it's trying to do ("match my_variable with either SomeValue x or SomeOtherValue y").
Another thing they make a lot of noise about is requiring fewer parens in some places. But then, they also require more parens in other places. So...okay? I guess? Not really a win.
And why rename equality operators? Are you really going to tell me that people prefer that their languages have "==="?