They’re merged and will be released in the upcoming GHC 9.6. So, we won’t have to wait that long any more.
Delimited continuations primops are about 1.5-2x as fast (compared to speff which seems to be the next fastest) in micro-benchmarks that mostly only test the overhead of the effect system itself. I’d still call that significant. Of course in many applications the effect system itself is not the bottleneck.
They use the same delimited continuations as eff, but then encoded as a “normal” Haskell data type. The implementation in that new paper is not that much different from the implementation described in this 2005 paper which is implemented in the CC-delcont package on Hackage. But I do agree that delimited continuations are more difficult to grok than the free monad.