“The O1 Pro and DeepThink models are theorized to operate by running the base O1 model multiple times in parallel and using a reward model to select the best output.”