The biggest value for me is:

- Really high SNR when searching about topics I am not a subject matter expert on.

- Really good at finding prior art on a topic while not necessarily limited to papers conferences and patents (e.g. finding obscure blog posts on a hobby project that actually is relevant prior work).

Arguably this can be achieved with frontier models but my experience is that Kagi is vastly more exhaustive and well I don't need to ask for references because they are indexed in the search itself.