Fine-tuned 4o errors using the responses API (with image input)

The first thing I set, knowing that it would loop, was the maximum output. Useless.

This fault will incur a huge bill, unstoppable by max_output_tokens. Here, setting it to just above the minimum value accepted (which reports the same 35k input usage that would have been 301t for the small image otherwise):

Varying the ranges of max_output_tokens doesn’t vary the bill of the same amount. Nor does terminating a streaming output. I imagine if you sent a message with many non-tiny images, the potential cost would not be limited at all by the model context, like you show. It probably also can re-bill for the “deny images” system message of 200+ tokens that is exempted, that OpenAI attempts to make secret.

This massively violates the API promise of max_output_tokens being a billing limit.

To stop any support time-wasting and render “provide a request id” useless to delay action:

• Empty response and nonstop output: x-request-id req_f48bdc55509442b6a033864f1d7aea8b, req_54a906c37e774230a84105c9b04f0caf
• Empty response and terminates on one: x-request-id req_e2422e2c709e496582b656864bd0ddf1

The playground seems to be only demonstrating one empty assistant output when it happens about 50% of the time currently, yet even at max_output_tokens=16 and the tiniest of images for 282t input:

Chat Completions:

1 in 10 or so, exactly the same fault as before there too, instead of going loopy:

Request ID for the Chat Completions API returning 400: x-request-id req_eafcb2667c344ec78350e71be2405bd3

The prior thread from two weeks ago with two weeks of inaction, then on gpt-4o, gpt-4.1 and gpt-4.1 mini fine tunes:

Fine-tuned Model Vision Capabilities Lost Bugs

Start requesting refunds for all the gpt-4.1 training credits in dollar figures - and it doesn’t have to be one trained on vision, because this makes any ft model unavailable for use in the image input pattern you want. Might get fixed.