“Systems like O1 Pro and O3 are distinct from earlier models because they can perform multi-sampling and recomposition of chain-of-thought reasoning at test time.”