Downvote the page, give feedback.
The change to “not subscriptable” is more the openai python library updates that came out at the same time as the additional features. The same pydantic model based openai library was available in beta test before any details of assistants, tools, or other API revelations were released.
I wrote a demo of getting out image data from the response format not documented. It uses ‘model dumps’ after slicing the object into metadata and the actual json shown.