with only 2 times difference, actual implementation of code generator also important
You cannot compare arbitrary code generators one with Kapt, one with KSP, because they actually do some work and performance depends on many factors, also build time increase is not not always linear with increase of amount of code/usages, it can be logarithimic or quadratic or worse
KSP is beneficial in terms of build speed for existing processors, because you do not get additional stabs generation step, but it doesn’t mean that you can assume performance of code generator only by the fact that it uses KSP
disclaimer: I know nothing about actual performance of Koin code generation and effect on build time, my comment is purely theoretical.
P.S. do not believe any performance claims without proper benchmark (and in many cases it’s very hard or impossible to have one) or until you measure it yourself on your own project with your use cases