But at the same time, the syntax itself is what matters to most of its users. OS thread can be fast enough for most workloads, but having the syntax showing up where the IO is happening, easily run concurrently what can be, and cancel stuff after a timeout, is the superpower of async/await.

Discussing about implantation performance is missing the forest for the tree. It's like comparing exceptions to error enums on that prism alone: sure there are some implementation where this or that is more efficient than the other, but the ability to express intent in the code is what make those killer features.