One of the assumptions flex makes is that interactive applications are inherently slow (they're waiting on a human after all). It has to do with how the scanner detects that it must be finished scanning a token. For interactive scanners, after scanning each character the current state is looked up in a table (essentially) to see whether there's a chance of another input character possibly extending the length of the match. If not, the scanner halts. For non-interactive scanners, the end-of-token test is much simpler, basically a compare with 0, so no memory bus cycles. Since the test occurs in the innermost scanning loop, one would like to make it go as fast as possible.
Still, it seems reasonable to allow the user to choose to trade off a bit of performance in this area to gain the corresponding flexibility. There might be another reason, though, why fast scanners don't support the interactive option.