chat completions is stateless — you must provide the entire conversation history with each new message; openai stores nothing (at least nothing that the downstream product _can use_) beyond the life of the request.

responses api, by contrast, is stateful — only send the latest message, and openai stores the conversation history, while keeping track of other details on behalf of the calling app, like parallel tool call states.

but i would say that since chat completions has become an informal industry standard, the responses api feels like an attempt by openai to break away from that shared interface, because it is so easy to swap out providers with nothing more than a base url and a model id, to a paradigm which requires data migration as well as replacement infrastructure (containers for code execution, for example).