Interesting findings. I believe dex2oat only slows down the installation time regardless of whether multidex is on or off. However, my investigation is more focused on the ongoing app startup time with multidex and not only on the first application launch.
“For that reason, first-run with multi-dex enabled is slower on dalvik vs. ART. Once dexopt has done its thing, loading from auxiliary dex files has the same performance as non-multidex.” is not fully true. We noticed the multidex performance impact from our automated tests which run app startup tests 100 times on each device. We found that there is no performance impact on the average startup time with multidex on 5.0+ devices, but devices prior to 5.0 slows down significantly. So the performance impact is not only on the “first-load”.