So I do not see them a sure win for projects I work on.
The only big F# feature I know of that's fallen out of favor is type providers, and, IMO, that's not because the feature was half-baked so much as because it was one of the many casualties of .NET's transition to being a truly cross-platform toolkit.
Do people just draw the comparison because they don't actually know of any other ML languages besides OCaml?
Things F# has that OCaml lacks: Extension methods, units of measure, type providers, quotations, computation expressions, active patterns, overloading.
Probably most of the F# things could be (and perhaps have been) added to OCaml with camlp4, which is more powerful than F#'s quotations. Versus, F# may never be able to have all of OCaml's ad-hoc polymorphism features. So there's that. But I'm mainly meaning to compare what's actually built into the language.
This won't always apply, but sometimes (many times?) there's a software endeavor that permanently feels on the margins of a programming language's capabilities, whatever that means. You may forego market share in order to be 5 or 10 years ahead of current practice. In other words, the mainstream isn't converging to a stable F# / Scala / whatever. It's lagging behind. It's not the last time that a mainstream language adopts a feature that has been available for years in a more fringe ecosystem.
In practice, I'm sure the most important thing is humans and their skill, so mainstreamness matters. But if you can somehow be less sensitive to that, you may reap a lot of benefit from using Foo before it was cool.