“OpenAI's internal research indicates that training a model to use one set of tools improves its ability to use other, different sets of tools, suggesting strong generalization in tool use.”