I’m using the “chat” in the playground to create an image generator. However, the image disappears immediately after being generated, leaving only the “Empty assistance message”.
I can reproduce this too.
I tested with gpt-5.5, gpt-4o, and gpt-5, using both gpt-image-1.5 and gpt-image-2 and I see the same behavior.
So what I understand, this does not look specific to one chat model or one image model. It looks more like a Playground UI/rendering issue where the image generation result is not displayed/persisted correctly and the chat surface may show empty assistant messages.
Although API Images works normally, so it’s just Playground Chat / Responses API that seems to have issues/bug.
Issue confirmed: the image is briefly shown, then obliterated.
Notable is that the “get code” doesn’t have the assistant response turn at all, despite that there are two “assistant” blank boxes in the UI and there should be event delivery output that is reflected in further inputs.
You edit text of one of the blank boxes, and then the assistant output_text is populated in code example, so it could be client-side code that is unaware of event handling when moving turns to “input” for another call.
You can control your costs and direct image creation better by using the ./images API, anyway.
I want to test some conservational image editing (multi-turn). In the playground, it seems that I can only do that via Response API.
Chat Completions is exclusively your own developer functions signaled as AI output. There are not OpenAI-hosted tools that run multi-turn autonomously on that endpoint.
Thus, playground is for simulating conversational image editing that you’ve programmed yourself (although the playground also cannot interleave images with text properly as you can do on the API).
For example: Here’s how you can test on the Chat Completions Playground that OpenAI has turned models into to hot garbage:
gpt-5.2, instead, for example:
or calling my custom image creator function, that might be based on the “edits” endpoint, but gives the user control:
You already got a great example above, but I’m just filling in with docs that may be useful for you and for others who have the same question.
Image generation guide:
https://developers.openai.com/api/docs/guides/image-generation
Migrate to Responses:
https://developers.openai.com/api/docs/guides/migrate-to-responses
From what I understand, if you want OpenAI-hosted image generation/editing in a conversational or multi-turn workflow, the Responses API is the better fit.
Hey @all! really appreciate you sharing this with us. We’re checking into it and will follow up here once we have more to share.
- Sunny