The spectrum will be identical for all cases of oversampling. (The frequency axis just scales accordingly). Proper oversampling does not modify the spectrum in band in any way whatsoever, but on the transmit side it provides greater margin to simplify subsequent filtering: either digital filtering if later interpolation / upsampling is performed, or the analog reconstruction filter after the DAC. Higher oversampling creates more spectral space between the images in the signal and therefore relaxes the transition bands in the related filters mentioned.
A possible benefit to lower sampling ratios is that the time duration of pulse-shaping filters implemented in the waveform creation (in this case the Gaussian filter) will extend longer in time for the same filter complexity and thus have higher performance. I detail that trade further at this post.
Therefore the choice of what oversampling is best is part of a system design with consideration of all subsequent filters needed, and traded with the resource and power requirements needed to run the digital processing at a higher rate.
Below is an example demonstrating a 16-QAM Spectrum and it oversampled by 5 where we can see from the details in the plot that the spectrum in band is not modified in any way, but the oversampling gives us much more range in frequency that facilitates later filtering operations both in the digital and analog domains.
16-QAM 4 samples per symbol:

16-QAM 20 samples per symbol:
