It does not appear that this aims to use an API, but rather to train/finetune their own model using insights from R1 ([mostly] pure RL approach towards bootstrapping reasoning in an LLM)