I used phenix to assign the free reflections putting in the twin operators. Doing simulated annealing in phenix, I get a rather large difference in R/Rfree of ~7%. Well, I guess I need to do some tweaking of the parameters in phenix (running the latest phenix and cci_apps). When I than use these free reflections assigned by phenix in Refmac I still get only 1.5% difference between R and Rfree?

Are you certain that you're using the same test set for both programs? Refmac expects that 0 marks the test set. Phenix will recognize this and adjust to use it automatically, but by default it still uses the Xplor/CNS convention where the test set is marked 1 and everything else is 0. If you flagged reflections with Phenix, it's possible that Refmac is using your test reflections as your working reflections. (I suspect that you'd see other problems too if this was the case, but I've made this same mistake before and I think it resulted in very close Rwork/Rfree.)

-Nat