Got to try GPT-5 a bit today (ChatGPT in o3 on some accounts seems to be routing
to GPT-5 already), it got better on Nim, which is good. First example:
https://play.nim-lang.org/#pasty=trnJEAWo mandelbrot in terminal with colors,
zoom and arrow movement. It's simple, but older OpenAI models couldn't really
make that work. GPT-5 did have some errors though: used "out" variable name,
used "UChar" instead of "cuchar" and forgot to import termios and exitprocs,
after fixing those it just works. And I was surprised to see that it actually
used the strformat features for formatting numbers properly, no model did that
before unprompted.

Second example is a Scheme REPL, it made some small errors, like forgetting to
"return" values properly in a "case" statement, putting "kind" twice in object
definition. There was also a small bug where it couldn't start, GPT-5 fixed it.
A fixed version is https://play.nim-lang.org/#pasty=WXuWVBpo There's still at
least one more bug (and maybe more) - multiline definitions don't work properly,
and there's no tail call optimization, but it's still impressive compared to how
previous models would fail at much simpler Nim things. Granted, most of that
1kloc file is boilerplate ;)

The fully released version should be much better when paired with an actual
IDE/tool (e.g. codex, claude code, cline) to let it see the errors and fix them
on the fly.