We present a $q_T$-resummed calculation of diphoton production at order N$^3$LL$^prime$+NNLO. To reach the primed level of accuracy we have implemented the recently published three-loop $mathcal{O}(alpha_s^3)$ virtual corrections in the $qbar{q}$ channel and the three-loop transverse momentum dependent beam functions and combined them with the existing infrastructure of CuTe-MCFM, a code performing resummation at order N$^3$LL. While the primed predictions are parametrically not more accurate, one typically observes that they are the dominant effect of the next order. We include in both the $qbar{q}$ and loop-induced $gg$ channel the hard contributions consistently together at order $alpha_s^3$ and find that the resummed $qbar{q}$ channel without matching stabilizes indeed. Due to large matching corrections and large contributions and uncertainties from the $gg$ channel, the overall improvements are small though. We furthermore study the effect of hybrid-cone photon isolation and hard-scale choice on our fully matched results to describe the ATLAS 8 TeV data and find that the hybrid-cone isolation destroys agreement at small $q_T$.