1326ms vs. 1382ms is 60ms, and when normal page load times are somewhere around 200ms, it makes a substantial difference.
Even if we disregard the fact that microtime tests have a big stddev, your test is not very convincing: You are using these constants several hundred thousand times whereas Drupal only has 2500+ in its *entire* source code. And of that source code, only a fraction gets to execution. I'd estimate the number of constants that are actually executed on a Drupal page view to about 500, and even that is probably too high. If you extrapolate your results (60ms for 100000), you'll get 0.3 msecs. Besides, a real world server behaves quite different from a desktop machine, especially in such edge cases. Konstantin