[kodapoda]

I use the free version and your example shows a vast improvement, indeed. 3.5 looks like a summary of Google results, while 4 actually gives a relevant answer. I don't have much insight into the differences between these versions, assumed that the biggest difference is "just" the training corpus.

[modeless]

It's also a much larger model which is slower and more expensive to run, which is why they can't offer it for free. The improvement is not always obvious at first glance but it is massive. I assume they will eventually be able to distill it into a smaller model that they can offer for free. I'm actually surprised that hasn't happened yet, especially given the fast progress in shrinking open source models.