If you're solely looking for accuracy, than I would say rhubarb. The precision of models of proteins embedded in a membrane is often arbitrarily high, especially considering the fairly low resolution structures commonly obtained for membrane proteins by X-ray crystallography, or the often under-defined structures yielded through NMR spectroscopy or cryo-EM. So in essence your model is only ever going to be as good as the template you model it on. However, I would define accurate as something which gives me meaningful insight into a biological system which I then can confirm through experimental data. Not all of that experimental data has to be generated in our own lab, it is often very useful to perform literature searches to see if somebody else's findings on the system agrees with prediction you derive from your model, i.e. you find some key residues interacting with other subunits, or lipids. Maybe someone else has identified certain mutations within those key residues that destabilize the protein or affect its function in other ways. 

As with any other non-membrane model it strongly depends on the resolution of the template, but also on the sequence homology between target and template. In general you would expect lower resolution in templates as Peter Teriete wrote above. Nevertheless it all depends on what question would you like to answer, if just general topology, specific contacts within the structure or molecular surface for docking or drug design purposes. I would consider that with a good homology accurate template you would be able to answer the first two questions. Also consider that if you´re interested in a specific site, and it is conserved, you can tolerate inaccuracy in other regions. Good luck!