First I would look at the models which are being chosen -- they may be very similar, and could even involve parameterizations which are not implemented in your downstream software you are using to build the tree (so, there might be a practical constraint influencing your selection).

Then, barring any other information regarding the fit of each model (which you haven't provided; I personally like hLRT, see Posada and Crandall 2001 attached), you have two different philosophies you could use: (1) take the most conservative (i.e., least parameter rich) model; or (2) take the most generalized (i.e., most parameter-rich) model. I would fall into the latter camp, in that I think if the signal from your data is very strongly in favor of e.g., a single transition and transvehrsion rate or not, then there is nothing about a more generalized (GTR etc) parameterization which excludes that. On the other hand, using a more conservative model when a more parameter-rich one may have been a better fit in theory, does constrain rates to those assumed by the model. Attached Abadi et al. 2019 for recommended reading.