There is no single answer for all users. There are pros and cons, which might apply to some folks, while not applying to other folks.
Simply having one size vs a different size really makes no difference, in terms of estimated time to win.
In that regard, it is 50/50. Your odds on winning are better with smaller plots. But that difference is not likely going to ever result in winning even one more block (unless you have multiple PB of storage).
Perhaps the advantage is 50.1 % / 49.9 %, in favor of k32.
But if you already plotted, then it is up to whether or not you want to spend the time re-plotting, and incur the additional electricity costs associated with re-plotting.
In my case, larger plots resulted in less plots, which fixed a problem I had with my harvester being overloaded. With the smaller plots, I exceeded some ceiling, causing my harvester to complain about 5+ second lookup times, and it was reporting the issue with numerous different drives – seemingly randomly. When my number of plots got reduced, that problem went away.
So less plots make for happier harvesters. But you probably have no harvester issues?
And for the long haul, larger plots sizes will not hurt. k32 will probably be fine, too. But long term survivability has a slight edge (if any) with larger size plots.
For new drives, I would go with larger plots.
For re-plotting, there are too many variables, most of which will never matter. Flip a coin.
Based on Chia docs I previously provided, and plot sizes we generate, assuming that the base plot has K value (e.g., -k 32):
any (K + i) value plot has exactly 2^i more proofs; no more, not less
any (K + i) value plot is (3 * i)% bigger, when comparing to i * K-value plots (thus higher k-value plot is less proof dense per TB); the overhead is not coming from the proof part, but rather from plot “headers” (for a lack of better word)
MM plotter is not producing “full” plots, rather drops some proofs, in exchange for the plotting speed (Max commented that it was a “substantial” speed improvment). The resulting plots are somewhat smaller, but by less than 1%. (This was the very first line of attack used by Chia to attack MM; however, as we know they backed off (maybe the last attack as well, as we farmers switched en masse to MM).)
Based on the above:
there is no ETW difference between 1x 2^i K value plots (e.g., 1 k34) vs 2^i * K value plot (e.g., 4* k32) (we can assume that those dropped proofs balance out for bigger sets)
by using (K + i) plots (3 * i)% of disk space is lost, thus for the same abount of space the ETW is longer by 3% for those higher k value plots (e.g., ~1 extra day for 1 month ETW for K+1; ~2 extra days for 1 month ETW for K+2 plot; therefore, k34 ETW will be ~24 days longer than k32 over a one-year period)
Also, as a corollary to plots getting bigger with higher K values, we can assume that regardless of how big drives we use, the worst case is that the free space left is just shy of one k32 plot (when plotting k32 only), i.e., less than 108 GB (again, using 10^4 rather than 1024^4). This implies that for any drive size at best we can replace 32 k32 plots with 16 k33, or 16 k32 plots with 4 k34 plots to keep the same number of proofs on the drive (again, if the free disk space left would be close to 108 GB, what is usually not the case)