I've actually settled on a very similar workflow - I mostly use Claude 4.6[1M] with adaptive reasoning disabled on High/Max for implementation, and then I'll do some combination of manual review in conjunction with GPT 5.5 xhigh.