Open source those models would enable major lawsuits for copyright violation. And I doubt that they own the rights to the training sets either way.

Open weight models are still useful to extract knowledge from. With much lower cost compared to hitting the API