Image Generation broken in all our custom GPTs since the new default model — the GPT returns the internal /mnt/data/…png path instead of rendering

GPT builders Plugins / Actions builders
,
1

Context: We maintain a production suite of custom GPTs (GPT Builder) used in a daily animation/storyboard workflow. All of our GPTs that rely on the native Image Generation tool broke at the same time, so this is a systemic builder-side issue, not one misconfigured GPT.

Summary
Since around June 9, 2026 (coinciding with the new default model rolling out to GPTs), every one of our custom GPTs stopped displaying generated images. Instead the GPT either:

  1. outputs the internal sandbox path as plain text, e.g. /mnt/data/msa_storyboard_clip10_4panel.png, with no image, or
  2. replies “the image generation tool is not available in this session.”

It worked reliably until that date. The same prompt produces the image correctly outside a custom GPT — so the model/tool works; it’s the behavior inside the GPT that regressed.

This is NOT an “enable the capability” issue — I isolated it. I tested two opposite Capabilities configurations on the same GPT:

  • (A) All four enabled: Web Search :white_check_mark:, Canvas :white_check_mark:, Image Generation :white_check_mark:, Code Interpreter & Data Analysis :white_check_mark:
  • (B) ONLY Image Generation enabled, everything else disabled: Web Search :cross_mark:, Canvas :cross_mark:, Code Interpreter :cross_mark:, Image Generation :white_check_mark:

It fails identically in both. Image Generation is on, by itself, and the GPT still won’t render — so there’s no capability conflict and nothing left to “turn on.”

This is unrelated to the Actions limitation. The image comes from the native Image Generation tool, not from a custom Action returning a file, so the known “Actions can’t return image/video files” behavior does not apply here.

Steps to reproduce

  1. Open a custom GPT with Image Generation enabled (reproduced with all capabilities on AND with only Image Generation on).
  2. Upload reference images and ask for a multi-panel storyboard sheet.
  3. The GPT returns /mnt/data/…png as text (or “tool not available”) instead of the image.

Expected: the generated image renders inline.
Actual: internal /mnt/data/… path as text, or a “tool not available in this session” refusal.

What I already tried (none fixed it)

  • Both capability configurations above.
  • An explicit GPT instruction telling it not to output /mnt/data/… and to render inline — helps ~50%, inconsistently (the model still reverts to the sandbox path).
  • Re-prompting (“don’t give me the path, render the image here”) — sometimes forces a render on a 2nd attempt, not reliably.
  • Brand-new chats, multiple accounts, multiple browsers/devices.

Evidence of links can be shared if enabled

2

I can confirm that I am seeing the same issue.

I have a valid ChatGPT Plus subscription, and image generation works normally in regular ChatGPT conversations on the same account. However, image generation does not work in any Custom GPT I create, even when the GPT is configured as minimally as possible and Image Generation is the only enabled capability.

I also asked friends and colleagues with their own separate ChatGPT Plus accounts to test the same behavior. They were able to reproduce the same issue on their accounts as well: image generation works in regular ChatGPT conversations, but does not work inside Custom GPTs.

I tested this with:

  • my original Custom GPT with Image Generation enabled
  • a completely new Custom GPT
  • a minimal neutral test GPT with no special instructions, no knowledge files, no actions, no web search, no canvas, no model selection, and only Image Generation enabled

Minimal test GPT setup:

Name: Image Test
Description: Test GPT
Instructions: “When the user asks for an image, create an image.”
Capabilities: Image Generation only
Knowledge files: none
Actions: none
Web Search: disabled
Canvas: disabled
Model: not specified

Test prompt:

“Create an image of a red ball on a white table.”

Result:

The Custom GPT does not generate an image.

The exact same prompt works correctly in a regular ChatGPT conversation outside of Custom GPTs.

I also tested this on multiple clients:

  • ChatGPT web app on a PC
  • ChatGPT mobile app on a smartphone

The issue occurs in both clients.

I have already tried disabling and re-enabling Image Generation, saving and republishing the GPT, creating a completely new GPT from scratch, using a minimal prompt, using a completely neutral GPT name and instruction set, and disabling all other capabilities except Image Generation. None of these steps changed the behavior.

This does not appear to be a prompt issue, an instruction issue, a client issue, or a general image-generation limit, because image generation works normally in regular ChatGPT conversations on the same Plus account and the same issue can be reproduced on several other independent Plus accounts.

It looks like the Image Generation capability is visible and enabled in the Custom GPT configuration, but the Custom GPT does not actually invoke or access the image generation tool.

For me, this is reproducible 100% of the time in Custom GPTs, while the same prompt works outside of Custom GPTs.

So I strongly suspect this is a Custom-GPT-specific tool-routing or feature-flag issue rather than an issue with an individual GPT configuration or account.

3

Adding another data point here.

I’m seeing a very similar issue with Custom GPT image generation.

In my case, I maintain several Custom GPTs used for visual design / commercial poster / key visual workflows. These GPTs were previously able to generate images normally, but since around the same period mentioned in this thread, they have become unreliable when using native Image Generation inside Custom GPTs.

The main failure modes I’m seeing are:

  1. The Custom GPT returns internal file paths instead of rendering the image inline, for example:

/mnt/data/kv_direction_1_dermatology_bestseller_podium.png
/mnt/data/kv_direction_2_blue_lab_routine_system.png
/mnt/data/kv_direction_3_fresh_skin_shield.png

  1. The GPT replies as if it cannot return the image directly in the current session, for example saying that it can only provide a clean render prompt instead.

  2. In some cases, the GPT appears to leak internal/tool-related reasoning or attempted tool invocation text instead of actually invoking Image Generation properly.

After more testing, I want to clarify an important detail:

At first I thought this might be specific to my own account, because the failure happens much more frequently for me. However, after asking friends to test the same Custom GPTs, they also experienced similar failures, just less consistently. For them, the GPTs may work sometimes and fail other times. For my own account, the issue is much more frequent and disruptive.

So this may not be a purely single-account issue, but the failure rate seems to vary by account/session.

This does not look like a normal prompt problem to me, because:

  • it happens across multiple Custom GPTs, not just one

  • it happens with different image-generation tasks

  • Image Generation is enabled in the GPT Builder

  • the issue can happen even when the prompt is simple

  • the same GPT may work in one account/session and fail in another

  • regular ChatGPT image generation outside of Custom GPTs appears more reliable than image generation inside Custom GPTs

The behavior looks more like a Custom GPT tool-routing / rendering / runtime issue. The generated image seems to exist internally in some cases, but the Custom GPT exposes the internal /mnt/data/...png path instead of displaying the actual image.

I have also submitted a support request, and it has been escalated to a support specialist. I’m adding this here because my symptoms match this thread closely, especially the internal /mnt/data/...png output behavior.

For me, the issue started around the same general timeframe discussed here and has continued until now.

It would be helpful if OpenAI could investigate this as a Custom GPT Image Generation runtime/tool-routing issue, especially because the failure rate seems inconsistent across accounts and sessions.

4

Hello to all newcomers in the forum.

I cannot say much except that I have seen the same phenomenon: either no images at all, or just a path.

But I can not tell more details, because I simply switched back to normal ChatGPT.

One workaround could be to enter the instruction protocol from the custom GPT at the beginning of a normal ChatGPT session, together with the instruction to run the session according to those instructions.

But of course this only works if it is your own custom GPT and other people do not use it.

5

Habe the same issues with CustomGPTs since 3-4 days. The Generation Skeleton appears but no Image at all. (Pro Account, Germany)