There are just so, so many problems with these benchmarks. (This comment from the author of Servant, a Haskell framework, nicely sums up a lot of the criticism: https://www.reddit.com/r/haskell/comments/6b3dlt/techempower...) But cherry picking the single query benchmark in particular makes the issue worse. Like... how many web apps do you stand up where you have one client you want to serve? At least the "multiple requests" benchmark is sort of plausible for the kind of workload you'd want to handle.

In most cases, though, you care a lot more about the 95th or even 99th percentile response times than you do the mean; I'd trade quite a bit percentage-wise off my mean response time if it brought the 95th percentile way down.

(This isn't to hate on PHP; PHP is fine. But you shouldn't choose it because the "single query" benchmark says it's fast.)