ChatGPT gen_id and file-service:// asset pointers explained: how to actually get your generated images

You dug into ChatGPT's conversation JSON — a data export, or the backend-api/conversation payload in DevTools — looking for your generated images. What you found instead was this:

{
  "content_type": "image_asset_pointer",
  "asset_pointer": "file-service://file-A1b2C3d4E5f6",
  "size_bytes": 1417238,
  "width": 1024,
  "height": 1024,
  "metadata": { "dalle": { "gen_id": "aBcD3fGh1jKl", "seed": 92014, "...": "..." } }
}

No .png, no https:// link. And every obvious way to turn those identifiers into pixels fails:

$ curl "file-service://file-A1b2C3d4E5f6"
curl: (1) Protocol "file-service" not supported

HTTP 404
<Error><Code>ResourceNotFound</Code>
<Message>The specified resource does not exist.</Message></Error>

Here is what each identifier actually is, why none of them download anything, and the routes that do work.

Why nothing in the JSON downloads

Ordered by how often each one is the wall people hit:

1. file-service:// is an internal URI scheme, not a network protocol. It's how ChatGPT's backend names a stored file internally. No HTTP client can resolve it — curl exits immediately with the Protocol "file-service" not supported error above, and a browser address bar just searches for the text. Only ChatGPT's own backend can exchange that pointer for a real link, inside your authenticated session. The scheme is an implementation detail, and internal details like this change without notice.
2. gen_id is generation metadata, not a file handle. It exists so the model can refer back to a previous generation inside the same conversation ("same scene, but at night"). There is no public endpoint anywhere that accepts a gen_id and returns an image — we tested the plausible candidates in ChatGPT image gen_id: how to retrieve and download generated images.
3. The URLs the app resolves pointers into are signed and short-lived. While you're using ChatGPT, the web app trades each asset_pointer for a signed download URL (a chatgpt.com/backend-api/... link that redirects to files.oaiusercontent.com/...). Those links carry timestamp + signature query params and are scoped to your session: strip the params and you get an HTTP 422 listing the missing fields; let them age and you get the HTTP 404 ResourceNotFound above. Full breakdown of those two failure modes: ChatGPT image gen_id URLs explained.
4. You're on the wrong side of a product boundary. Everything in the conversation JSON — pointer scheme, metadata shape, resolver endpoints — serves the ChatGPT app, not integrators. None of it is a contract. That's not a bug to work around; it's the signal to use an actual API for anything programmatic.

How to actually get your images

1. One image, right now: download it in the UI. Open the conversation and use the download button on the image. This is the supported path, and the file you save is yours to host anywhere.
2. Your whole history: request a data export. Settings → Data controls → Export data. The zip contains conversations.json plus asset files named by file id, so file-service://file-A1b2C3d4E5f6 matches a file-A1b2C3d4E5f6-*.png in the archive. Caveat: exports cover what's still stored — images from old conversations whose assets have already expired server-side can be missing.
3. Only have a gen_id? Regenerate instead of excavating. A gen_id cannot be converted into pixels after the fact. If the image mattered, reproduce it from the prompt — and if this keeps happening, that's the sign your workflow belongs on an API.
4. Anything automated: generate through an API contract. On hiapi, gpt-image-2 — the API line of the model behind ChatGPT's image generation — is available through the unified task endpoint POST /v1/tasks (see the gpt-image-2 model page). The task result hands you a downloadable output[0].url with an explicit expireAt — a documented pickup window instead of a reverse-engineered one. The standard text-to-image line is $0.03 per image at 1K resolution (live pricing).
5. Re-host immediately, everywhere. This rule is universal — ChatGPT exports, hiapi output URLs, any provider's signed links: download the bytes, upload them to storage you control, and store your URL. Signed links that expire are the industry norm; permanence is something you own, not something you fetch.

Minimal working example

Create a task with the bare model id (gpt-image-2 — appending pipeline suffixes like /text-to-image to the id returns a 400):

curl -X POST https://api.hiapi.ai/v1/tasks \
  -H "Authorization: Bearer $HIAPI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "input": {"prompt": "a paper crane on a desk, morning light", "aspect_ratio": "1:1", "resolution": "1K"}
  }'

Success returns {"code":200,"data":{"taskId":"..."}}. If you instead get HTTP 401 with:

{"error":{"code":"permission_denied","type":"hiapi_error","request_id":"..."}}

your key is missing, wrong, or not enabled for the model — see the authentication docs and check your key in the dashboard.

Then poll the task and download the output as soon as it succeeds:

curl -s https://api.hiapi.ai/v1/tasks/$TASK_ID -H "Authorization: Bearer $HIAPI_KEY"
# when "status":"success": download output[0].url before its expireAt
curl -s -o crane.png "<output[0].url>"

Notice the contrast with the JSON archaeology above: the URL is part of the response contract, the expiry is an explicit field, and nothing depends on a browser session. (Yes, hiapi's output URL expires too — that's deliberate, and the fix is the same re-host habit: download before expireAt.)

Related reading

• ChatGPT image gen_id URLs explained: why backend-api links don't work
• ChatGPT image gen_id: how to retrieve and download generated images
• gpt-image-2 on hiapi: capabilities, pricing, playground

FAQ

What is asset_pointer in ChatGPT's conversation JSON? An internal reference to a stored file, written as a file-service://file-... URI. The ChatGPT app exchanges it for a temporary signed download URL inside your session. Outside that session it resolves to nothing.

Can I convert a gen_id into an image URL? No. gen_id identifies a generation for in-conversation follow-ups (edits, variations). No public endpoint resolves it to a file, and it isn't a parameter on any image API.

Why does the files.oaiusercontent.com link from my export/JSON return 404 ResourceNotFound? It's a signed URL that has expired or lost its query parameters. Those links authenticate with embedded timestamp + signature params and are meant for immediate, in-session use only.

Does the ChatGPT data export include generated images? Generally yes for recent conversations — asset files sit in the zip named by file id, matching the asset_pointer values in conversations.json. Assets that already expired server-side can be absent from the export.

Is there an API that accepts a gen_id or asset pointer? No. For programmatic image generation, call an image API directly. On hiapi that's POST /v1/tasks with model gpt-image-2: submit a prompt, poll the task, download output[0].url before expireAt.

How do I get a permanent URL for a generated image? Host it yourself. Download the bytes (from the ChatGPT UI, the export, or an API output URL) and upload them to your own bucket or CDN. Every provider-side URL — ChatGPT's or an API's — is a pickup token, not hosting.