I also noticed that in the docs for batch, it only lists 3 endpoints that can be used, and /v1/fine_tuning/jobs is not among them.
I assume I am missing something. (Maybe this is a feature available to only certain users? Maybe I completely missed an important part of the docs - not at all unlikely).
Fine-tuning already has a quality of not getting the model trained until openai gets around to it.
It is not available and doesn’t make much sense to “batch”.
A parameter that the fine-tuning job could, like batch, similarly be scheduled for off hours with low priority would be something that OpenAI could offer, but they don’t.
Let’s say you want to use chat completions and gpt-4o-mini to give you some metadata search keywords for each of 1000 web pages on your site.
With batch, you don’t have to make 1000 API calls. You can put the requests all in one file. You get the discount for a 24 hour turnaround where OpenAI can perform the calls during their low utilization times (or sometimes, cancellation after 24 hours, it seems).
The fine tuning infrastructure and servers may be a different resource, and already queue the jobs. That, plus the fact that you are going to fine-tune anyways, may be why there is no discountable opportunity yet given.
The pricing is when you use that model on chat completions for inference. Using your fine-tune through the batch API.
If you think differently, an industrious person could put their JSON API call to fine tune into a batch file, set the endpoint, and try to submit it to see if the pricing people know something that the documentation authors don’t.