I have compiled both versions with optimizations on, looks like the gap gets a bit smaller (percentage-wise) after that:

       benchmark                    |  master (ns)  |  PR 15633 (ns)  |  Δ (%)
------------------------------------+---------------+-----------------+----------
call only (datetime)                |     73 (±3)   |     92.3 (±7)   |   26
constructor + call (datetime)       |    228 (±9)   |     260 (±16)   |   14
timedelta + call (datetime)         |    108 (±5)   |     128 (±9)    |   18

If this were something fundamental like a performance regression in building a tuple or constructing a dictionary or something I'd be concerned, but this just reinforces my feeling that, on balance, this is worth it, and that we are probably not going to need a "fast path" version of this.