According to the screenshots of XMRig for android you only get about ~35H/s while my laptop does ~2400. That's 68x faster so if it took my laptop 2 seconds it would take a mobile device ~130 seconds.

It screws with mobile users and makes the whole crypto PoW thing about it using too much energy many times worse. Not to mention botnets could make use of enough computing power to easily outpace any captchas thrown at it.

Is that for the specific proof of work algorithm mCaptcha uses? While I don't think you're going to get something that runs equally quickly on a low-end phone and high-end desktop, if it depends entirely on sequential operations and is not optimization-friendly you should be able to get much closer than 68x?

I think we‘d need to compare apples to apples, and not use Monero mining as a benchmark for mCaptcha. Also, as I wrote in another comment, the average case (server is not under attack) is 0.02 seconds on a laptop, and probably 0.4s on an Android device even if we do use xmrig-android as comparison. Compare that to manually identifying stairs on pictures with crappy quality (10 seconds?).