Most major companies have released new versions in the past year, but if people were asked to blindly determine whether they were using the newer or older version, I suspect the results would be close to random. It seems to me that the difference between versions is sharply decreasing in a way that seems to be asymptotic, similar to what happens in literally every other domain with neural networks.
I also think it's clear that the difference between the various choices is also diminishing. Aside from certain manually designed idiosyncrasies (like ChatGPT's obsequiousness), I think people assessing which model they're using would also be mostly random. Somewhat surprisingly, even in the 'LLM arena' [1], where you get to compare output side by side, the difference between models is approaching statistical 0!
[1] - https://huggingface.co/spaces/lmarena-ai/lmarena-leaderboard