That's not the right questions.

If you have $500k to build a GPU cluster then you can get vastly superior performance than any of the regular subscription based AI solutions.

The limit is how much you can spend on hardware, not the models. It is about selecting the best model for your hardware and purpose.

You also need to download and try out the latest models as they come out, since the "best" open source model changes every month or so.