Custom GPT stopped generating images and returned `/mnt/data` file paths instead

GPT builders Plugins / Actions builders
1

Title

Custom GPT stopped generating images and returned /mnt/data file paths instead

Post

I have a custom GPT that I use for marketing content and image generation. It used to generate images normally, but recently it began creating PNG files and responding with internal file paths such as:

/mnt/data/image_name.png

The path was not clickable, and the image did not appear in the conversation.

After troubleshooting using ChatGPT, I found the problem was a conflict in my GPT instructions. The instructions were to preserve uploaded logos and other brand assets exactly, and I also allowed it to use code for precise placement of those assets.

When I requested an image that included an uploaded logo, the GPT interpreted the task as file compositing. It used Code Interpreter or another file creation method instead of the built in Image Generation capability. It then returned the internal file path rather than displaying the image.

I was able to fix it by replacing the image instructions with explicit routing rules (also with the help of ChatGPT):

For every request to create, generate, design, draw, visualize, or edit an image, always use the built in Image Generation capability.

Never use Code Interpreter, Data Analysis, Python, HTML, SVG, or another file creation method to make or assemble an image unless the user explicitly asks for a code based graphic.

Do not answer an image request with a prompt, description, code, filename, raw path, temporary path, internal path, or sandbox path. The finished image must appear directly in the conversation.

Once that was done, the image generation started working again.

I’m sharing this because it seems to be a reasonably common issue and I’m hoping it can save someone else a lot of headache

2

Thanks for sharing the detailed troubleshooting and workaround, @LCNCJohn. This should be helpful for others who run into the same /mnt/data behavior, especially when custom instructions encourage code-based image compositing.

There was also a recent issue affecting Image Generation that was posted on the OpenAI Status page. That incident has now been marked as resolved, so it may have contributed to what you were seeing alongside the instruction conflict you identified.

Are images now appearing normally in the conversation, or are you still receiving internal file paths after the incident was resolved and the routing instructions were updated?

-Mark G.

3

Thank you so much, LCNCJohn! I can’t begin to tell you how much time I wasted trying to make this work.

4

Yes! Images appear normally in the conversation now. They always worked outside of the custom GPT but stopped working in it.

Fortunately, ChatGPT is pretty good at auditing itself so I pasted my instruction set into it and asked why that was happening and had it solved in about 15 minutes.

5

I can reproduce this independently, and the instruction workaround did not resolve it.

The problem occurs only inside custom GPTs. Across multiple custom GPTs, when any Thinking model or mode is selected, ChatGPT generates a PNG but returns only an internal path such as /mnt/data/0.png; the image is not rendered in the conversation. Instant mode works correctly.

I reproduced it in new conversations on the ChatGPT web app, Windows app, and mobile app. Image Generation is enabled, and I tested after updating the GPT instructions to force the built-in Image Generation capability and prohibit Code Interpreter or file creation.

This appears to be a custom-GPT image-rendering or tool-routing issue affecting Thinking mode across platforms, rather than an instruction conflict in one GPT. My plan is ChatGPT Pro.

6

Thank you, LCNCJohn, for documenting the workaround, and thank you to OpenAI Support for clarifying the possible instruction conflict. In my case, however, the instruction-based fix did not resolve the problem.

I have now reported the issue through the official OpenAI Help Center Support chat and provided the requested reproduction details. My final findings are:

• Scope: the problem occurs only inside custom GPTs and can be reproduced across more than one custom GPT. It is not limited to one GPT’s instructions.

• Models: GPT-5.6 Thinking fails across the Thinking modes/levels I tested. GPT-5.5 Instant works correctly in the same custom GPT.

• Platforms: I reproduced the same behavior in the ChatGPT web app, Windows app, mobile app, and multiple browsers, including new conversations.

• Configuration and troubleshooting: Image Generation is enabled. Code Interpreter/Data Analysis was disabled, and I explicitly instructed the GPT not to use code, Python, HTML, SVG, or file creation for image requests. The issue continued. The linked reproduction does not depend on Actions or Knowledge files.

• Exact reproduction: using the prompt “remaster,” the custom GPT runs for 47 seconds and then shows:

Worked for 47s

/mnt/data/remastered_portrait_9x16.png

The generated image is not rendered in the conversation; only the internal sandbox path is exposed.

Failing conversation:

https://chatgpt.com/g/g-69cf1dd7cf4c8191a0200daf4ad73c35-raw-image-generator/c/6a7028a7-3ac8-83e9-9f50-5340f47e5f44

The current workaround is to use GPT-5.5 Instant for image generation in custom GPTs. Based on these tests, my case appears to be a custom-GPT image tool-routing or rendering problem affecting Thinking mode across clients, rather than an instruction conflict in one GPT.

I hope this additional evidence helps the developers distinguish this case from the resolved incident and investigate the Thinking-mode custom-GPT path separately. Thank you again to everyone who has taken the time to investigate and share information.

7

Thanks for reporting this, @LCNCJohn. Also appreciate @Skibee and @edwardvonfuchsberg adding their experiences and helping confirm the broader issue.

After further investigation, we’ve confirmed this is a bug and the team is working on a fix. For now, the workaround is to use GPT-5.5 Instant, regular ChatGPT, or Projects for image requests instead of GPT-5.5 Thinking or GPT-5.6 Thinking inside a Custom GPT.

Thanks again to everyone who shared details and workarounds.

-Mark G.

8

I have exaclty same issues that other people have. I think i was alone but this was frustrating before i find this topic. Please OpenAi fix this because i love to use GPT with the 5.6 thinking model. And correct it works only with 5.5 Instant model…

9

Welcome to the Community, @mackan3394. I can see how frustrating this has been, especially before finding others with the same issue.

Our team is actively investigating this and working toward a fix. Thanks for your understanding.

-Mark G.