The all-electron CCSD(T)/aug-cc-pCVnZ relative energies differ from the complete-basis-set limit by about 1000, 300, 100, and 50 cm(-1) for n = D, T, Q and 5, respectively.We test the accuracy of various standard, explicitly correlated F12, and composite ab initio methods with different correlation consistent basis sets for high-dimensional potential energy surface (PES) developments, thereby providing a practical guidance for reaction dynamics studies. Relative potential energies are computed at IS geometries covering the energy range and configuration space of chemical importance for each of the six prototypical polyatomic reactions, X + CH4 -> HX + CH3 [X = F, O, Cl] and X- + CH3Y -> Y- + CH3X [X/Y = F/F, OH/F, F/Cl]. The average accuracies of the Hartree-Fock and MP2 methods are 1500-8000 and 400-1000 cm(-1), respectively. The standard CCSD(T) method provides errors of 900-1400 and 250-450 cm(-1) with aug-cc-pVDZ and aug-cc-pVTZ basis sets, respectively. The explicitly correlated CCSD(T)-F12 method reduces the corresponding errors to about 200 and 100 cm(-1); thus, we recommend using the F12 methods for PES developments. For F12 computations, the cc-pVnZ-F12 [n = D and T] basis sets usually, but not always, perform better than the corresponding aug-cc-pVnZ bases. We do not find clear preference between the F12a and F12b methods for PES developments. Composite methods are advocated instead of standard CCSD(T) because for example, one can obtain CCSD(T)/aug-cc-pVnZ quality results on the expense of MP2/aug-cc-pVnZ [n = T and Q] computations. The post-CCSD(T), the core correlation, and the scalar relativistic effects are found to be similar to 100, 80-130, and 10-50 cm(-1), respectively.