Didn't the CEO say during the GPT-4 demos that humans were involved to do the captions during all the images tests (using Be My Eyes volunteers) and that's why it took some time to show the Discord examples ?
If so, it could explain some costs indirectly.