Because it is built on the same algorithm (not just based on the same theory, but the exact same algorithmic code) as the real test, and because it uses real past GMAT questions, it is definitely the most accurate.

This doesn't mean it's incredible precise, however. The official test has a standard deviation of 30 points, which means that you're "likely" score is considered to be in a 60-point range around whatever you score. (Eg, if you score a 650 on the real test, you're "likely" range were you to take the test again would be 620-680). So it's not the case that, if you get a 650 on GMATPrep, you should expect exactly a 650 (or a 640 or 660) on the real thing. In fact, substitute "real test" in place of "GMATPrep" in the above sentence, and the sentence is still true. You're considered just as likely to score a 620 or a 680 if you take it again.

Kind of annoying, I know - but you do have to factor in that idea of standard deviation - the test is not anywhere near as precise as most people assume it is.