The answer is pretty obviously AV1 if you care about delivering content to actual end-users, unless you expect them to be
severely bandwidth constrained such that the ~30% improved efficiency makes a noticeable difference. AV2 is still not mature from an implementation perspective; I wouldn't consider it ready until at least the dav2d software decoder has had time to mature, which the project page implies is still in a fairly early stage of development (they still list "complete C implementation of the decoder" as "on-going":
https://www-test.videolan.org/projects/dav2d/) and I am not sure if any encoder exists besides the AOM reference implementation (which was notoriously slow for AV1, so I assume the same for AV2). Furthermore it will be years before hardware AV2 decoders and encoders make their way into consumer products (this will likely be further delayed with the current shortages of memory and storage chips causing hardware lifecycles to be extended even more than they already had been over the past decade).