Switching from test to test in the server takes more time than your code switching from case to case, and it reduces the number of verdicts.

It makes the waiting time to get the final verdict smaller.

See reply to this comment.

Apart from aforementioned testing speed-up reason, for many problems only a small percent of inputs can fail the wrong solution (for example, when greedy approach is almost optimal). And even T=100 effectively increases the probability that the wrong solution will fail the pretests.

There are a yes-no question and 10 pretests to it. The greedy would pass, and a lucky man would get pretest passed by random solution, and abuse the system by submitting it by another account, try hacking, and copying the other's code for the real solution.