"Random guesses" is probably underselling the capabilities too much, given their internal tests claim 0.0041% false positive rate. Obviously their tests might be biased and designed to make them look maximally good, but at the same time it's more implausible to claim they do no better at random guesses.