His timelines show zero multicam. I have edited small amounts of single-camera 4k H264 on my 2013 top-spec MacBook Air, without proxies. It is probably assisted by the low screen resolution. The previous MBA was 1440x900, vs the 2018 model at 2560 x 1600, roughly 3x the pixels, and the same as his 2016 13" MBP.
For more substantial amounts of multicam 4k H264, I don't even like using my top-spec 2016 MacBook Pro. From his prior videos it appears he has a 2016 top-spec 13" MBP.
He said he used "better performance", which means he wasn't using proxies or optimized media.
It appears starting with Kaby Lake, either Quick Sync or how FCPX uses it was greatly improved. My 2017 i7 iMac is about 2x faster on scrubbing and exporting 4k H264 than my otherwise identical 2015 i7 iMac. I've always wanted to do an H264 FCPX performance test on the Skylake 2016 vs Kaby Lake 2017 MBP to see if there's a similar improvement, but never have. According to AnnandTech, the Kaby Lake Quick Sync AVC (meaning H264) encoding performance received a boost over Skylake: https://www.anandtech.com/show/10610/intel-announces-7th-gen-kaby-lake-14nm...
I'd expect this to help the 2018 MBA compensate for its slower clock rate, at least when decoding/encoding H264.
Of course this only helps if both application and system software layers support this. Premiere (even 2019) apparently does not use Quick Sync for decoding, only encoding. FCPX and Resolve also use it for decoding which may explain why timeline and other scrubbing operations are so fast on those vs. Premiere. The implication is (apart from exporting) I'd be hesitant to expect his good FCPX results on the 2018 MBA if you are using Premiere.
It's interesting his export test was about 2x faster on his 2016 dual-core 13" MBP. His MBP is a 3.3Ghz i7 model with 16GB RAM, which had integrated IRIS graphics 550. The IRIS graphics units don't have any VRAM but allocate up to 1.5GB of system RAM. The higher-end IRIS units have an "eDRAM cache" to minimize contention for system RAM. His 2018 MBA gas a 1.6Ghz dual-core i5 with integrated Intel UHD graphics 617, which has no eDRAM cache.
Since he's comparing two dual-core laptops, maybe the higher clock rate alone helps explain the approx. 2x better encode time on the 2016 13" MBP. His timeline showed no render dots so it appeared both were fully rendered before export. If it was fully rendered (meaning no effects to compute), it should mostly be an encode test. I'm actually surprised the MBA wasn't closer to the 2016 MBP on that one test.
I wonder how much more popular the 2018 MBA would be if it had an SD card slot?