The time to spawn the threads is just inconsequential compared to the time to wait for the threads to exit for small sizes of N. But by doubling the number of numbers to sort you do double the time (actual wall clock time) that it takes to spawn those threads. O(N + 10000000) is still O(N)

Not to mention for the OS to handle N threads would likely take some kind of non-linear time, at least N^2 if I had to take a shot in the dark. But I imagine it'd depend on the OS and how the hardware handles interrupts. Even if the thread is sleeping it still takes resources to schedule it