Based on uops.info these should have 5 cycle latency as they did on Haswell/Broadwell. I have no additional internal information from Intel.
This was also shown as a discrepancy in the spreadsheet that was sent with an early llvm-dev post about llvm-exegesis.