One simple thing you can try is to go with the default maximum allowed error rate (--p-error-rate) instead of increasing it, as you have done. And if you haven't done this already, I would also suggest taking a deeper look at cutadapt's documentation regarding the minimum overlap parameter to make sure you're choosing a number there that makes sense.
Let's see how that goes. Not sure how big of a difference this will make. There is some more relevant discussion here as well: