Additionally running GPT-5.5 on medium sometimes gives me better results than high mode. On any of them I still have to push the models in the right direction.