Okay, USER ERROR!
I was using the wrong model 
This works: gpt-4-1106-preview.
This doesn’t: gpt-4-vision-preview.
So, note that you can’t get JSON output with the vision model, at least with the current preview.
2 Likes