This isn't adding full support for 3.12 (by advertising via a Trove classifier or distributing wheels), which is only just in alpha, but lets us test early to find issues in our code or even in CPython itself. Full release is October 2023.

Python 3.7 is still supported by CPython until 2023-06-27.

However, scientific Python projects following NEP 29 have been dropping 3.7.

Here's the pip installs for ujson from PyPI for October 2022:

Source: pip install -U pypistats && pypistats python_minor ujson --last-month

That's a pretty big percentage (but NumPy is similarly has 3.7 top with 38%).

I think we should follow either CPython's life cycle or NEP 29, to make things clear for users so they can plan ahead.

Generally scientific projects follow NEP 29 and non-scientific follow CPython, but there's a lot of variation.

Any preference?

No strong preference from me, I would tend towards CPython as this is virtually all C code and mainly affects production of wheels, but don't mind too much either way.

I can't think of any 3.8 feature that would be necessary or useful here currently. I'm not aware of any significant C API changes, and changes at the Python layer (e.g. assignment expressions) aren't relevant except for testing. So I don't see any reason not to keep 3.7 support until it reaches EOL next year.