Hi all,
I wanted to surface what seems like a recent and potentially breaking change when using DALL·E-generated images in openaiFileIdRefs
for POST requests or GPT Actions.
The issue
When I generate an image with DALL·E and immediately try to use its file-xxxxx
ID in a POST request via openaiFileIdRefs
, the request fails with the following error:
`GetDownloadLinkError Encountered exception: <class 'file_service_client.client.GetDownloadLinkError'>`
This seems to happen even before the file is passed to an external API—so somewhere during OpenAI’s internal expansion of the file_id
into a full JSON object (with download_link
).
This used to work flawlessly
According to OpenAI’s documentation, openaiFileIdRefs
supports DALL·E-generated images. And indeed, I’ve used this exact setup dozens if not over a hundred times over the past several months without a single issue.
But starting around two weeks ago—coincidentally right after a DALL·E model update—this same approach started failing consistently.
Repro steps & what I tested
-
Used a fresh DALL·E image
-
Captured its
file-xxxxx
ID -
Added it to
openaiFileIdRefs
-
Attempted POST action immediately (and even after waiting 10+ seconds)
-
User-uploaded images still work fine
-
DALL·E file IDs now fail 100% of the time
Even delaying the request 10–20 seconds after generation doesn’t help, unless I explicitly download and re-upload the image file myself.
This seems to indicate that the image is not fully “registered” in file-store or exists only ephemerally.
Why this matters
-
This behavior contradicts OpenAI’s docs, which don’t warn of any such limitation.
-
It breaks workflows that have worked reliably for months.
-
It’s not just a delay issue—it seems systemic, and likely related to recent backend changes.
Anyone else seeing this?
I’m curious if other developers here are running into the same issue.
Is this a recent change to DALL·E or file-store behavior?
Or worse, an unintended regression caused by recent infrastructure updates?
If it’s a bug…
If this is a bug, I sincerely hope OpenAI picks it up quickly—this effectively makes DALL·E output unusable in automated workflows unless we manually re-upload each image, which defeats the whole point.
Thanks in advance to anyone who can confirm or offer a workaround. Would love to know if this is on OpenAI’s radar.