The estimate in the PEP is that it will be 5-8% slower. Having to use more granular locks and atomic operations has a cost.

https://peps.python.org/pep-0703/#performance