Don't they need access to the models to use them for Bing?

I would consider those models "published." The models I had in mind are the first attempts at training GPT5, possibly the model trained without mention of consciousness and the rest of the safety work.

There is also all the questions for RLHF, and the pipelines to think around that.

Not necessarily, it would be just RAG, the use the standard Bing search engine to retrieve top K candidates, and pass those to OpenAI API in a prompt.