These benchmarks are all flawed.
You can't measure input latency properly without a camera (or pixel access to the screen but then have to be careful it's not impacting your benchmark). The latency benchmark is just testing VT IO throughput for a specific zsh prompt. Not at all related to input latency.
The plaintext IO benchmark is suspiciously slow for both Alacritty and Ghostty relative to Kitty. I ran the same on my machine and got within 15% between all three. They're all fast but something looks wrong here. I also have issue with the contents, which aren't representative in my opinion of human output (that's why I like to use books as my plaintext IO benchmark). This changes things.
The memory usage doesn't explain the setup at all, nor exactly how memory is read (I'm not sure what the `mem` column is reading in btop). Memory usage on macOS is tricky to benchmark because the OS tries very hard to _not free memory_. As a macOS kernel guy told me once: "freed memory is wasted memory" [so long as you're not using all memory]. There are other metrics you can get out of the kernel to get accurate _allocated and used memory_. Activity Monitor exposes these well. (EDIT: Just realized the tests were on Linux, which should be better, so this might be okay but test setup still matters here)
I'm not saying any are directionally wrong, and I'm also not trying to claim Ghostty is the fastest (I don't, I claim it's just "fast", not the fastest). But this is not good benchmarking.