This is a persistent myth that’s only true if your performance model requires persistent latency and, even then, only if you’re careful to pay attention to how much is freed or allocated as you change scopes. The lifetime tracking features of Rust’s type system have non-negligible impacts on the sorts of abstractions you can make.
Sure, there is one abstraction that effectively requires GC, namely a general graph of spaghetti references where you can't possibly predict in advance or "pay attention" to" how much is freed or allocated as you change scopes". Many GOFAI problems look like that which is why GC was initially developed in the context of LISP, the most prominent language for GOFAI. But for most real-world programs you can do vastly better than that.
