There was a point in time recently, when you could get the image with a transparent background on a platform like you recreated - that texting the image to yourself provided the transparent version without filling in the transparent bits with a solid checkerboard.
Bleatlock would know more about that.
I did get that method to work, it just ended up being quicker for me to use the solid green chroma background…
Looking forward to finding out if this is just a bug or something openAi is walking away from supporting?
It is not a bug. OAI docs says that gpt-image-2 doesn’t currently support transparent backgrounds
My guess is that gpt-image-2 was optimized for other priorities first, higher realism, arbitrary resolutions, better generation/editing quality, maybe speed/architecture changes and transparent alpha output did not make it into this version. (This is just a guess based on docs)
Hmm, why do think that? I’m not sure I understand Because transparent backgrounds are still supported by previous image models. I really think, like I mentioned above, that there might have been other priorities of capacities for this version
I don’t have video/MP4 samples to share. My testing here was only for still-image transparency.
For video, transparency depends on the codec/container workflow supporting an alpha channel, so I wouldn’t assume ordinary MP4 output has the same behavior.
All parameters with a question mark are optional - and you still get an image with an empty JSON as merely a trigger (and contents by the AI model observing the conversation).
Notable is that there are…
Leaving the AI to guess using pretraining up to cutoff or post-training on tool calling, and to hallucinate that referenced image IDs still does anything or that more than one image could be delivered from this old tool spec.
So we specify. Notably, ChatGPT gives us no promise of what “model” is used, or if today you fall back to a cheap “mini” image creator that does support transparency, or a “beta”. What we do know is that gpt-image-1.5 that is the last API model to offer transparency, and it is limited to three sizes. So we specify, like the earlier poster’s insight
Notable is that there is no “n:4” being satisfied, but instead the image model that also views the entirety of the chat interprets the language I sent (and not the internal tool JSON) as making a grid of four.
However, success: the images are indeed transparent without any “post-coding”. Here they are in my application where this mustard color rarely seen is the background behind transparency.
The delivered and seen whites are a value about 252 of 255, which aligns with prior suspicions of transparency being a held-out brilliant white, programmatically removed by OpenAI (and prior images being that sepia filter color).
I did a little easy test in ChatGPT and gen actually did surprised me a bit with the steps to transparent background. And this time fairly easy. At the beginning I prompted “fox cub trying to reach for a butterfly, orange background”, then the rest of the steps can be seen in screenshots
Usually when I’ve tested with opaque color and then edited to true alpha channel, it results in almost transparent background. But this time, chessboard and then with a simple prompt “remove the chessboard background” and the output was completely transparent.
Add this paragraph when generating the image, it works for me all the time
“Make sure the image uses an alpha channel with a fully transparent background. No white, gray, checkerboard, or colored background. The final image should be a clean PNG-style logo with no background at all”
I tried the same prompt as yours, and here’s the result. I’m not sure why it’s not working on your side, but one thing worth mentioning is that I’m using the “Extended” Thinking model, not the standard one.
GhatGPT 5.5 Thinking, seemed to have generated a fake transparent background for me. Tho using my phone, it shows as transparent. I have by default white screen and when I press the image it changes to black background and the image seems to have transparent background.
Because transparent backgrounds are still supported by previous image models. I really think, like I mentioned above, that there might have been other priorities of capacities for this version
and great prompt 
