I believe I heard that running inference longer is giving the better responses we're seeing in v4. Hence why v4 is taking so much longer to output data.

Of course we won't know this for sure until OAI tells us, so we may be in the dark for a while.