This isn't really correct. In many (most?) multithreaded applications a thread will regularly yield, forcing a context switch, long before its kernel time slice expires. In modern systems it is not difficult to have a case where the time spent doing work is less than the time spent yielding to another thread i.e. almost all of the CPU time is spent context switching.

This is in significant part a side effect of CPUs becoming much faster. You need to increase the amount of code that is run between potential thread yield points to offset the overhead of a yield or it can generate context switch storms (i.e. almost all time is spent context switching) and throughput will suffer.

Modern server architectures take this to its logical conclusion and try to eliminate virtually all context switching, hence the rising prevalence of "process-per-core" software designs.