Following image is from a custom GPT.
Today, I received several times a message in some custom GPTs like below.
It looks custom GPTs will use 4o Image Generation soon.
Made with the old version of image generation. New images coming soon.
and
This image was created using DALL·E, OpenAI’s legacy image generation model. A newer model is rolling out soon within ChatGPT, so stay tuned for even more enhanced visuals!
transform this image into a 3d model on a display table, covered by a rectangular glass protection
A wide aspect ratio photorealistic, slightly shaky WhatsApp video call screenshot, captured in a dimly lit, mysterious tech-lab atmosphere. The screen shows Sam Altman (late 30s, sitting in front of glowing whiteboards. Behind him are scribbled notes like “VIRAL AGENCY THEORY” and “4o core protocols”, along with neural diagrams and complex equations.
The UI is accurate to WhatsApp with green tones, end-to-end encryption notice at the top, and a visible call timer showing “2:00:17”. The time displayed in the upper corner is “03:14 AM”, with a critically low battery icon at 3%. Signal bars are poor. The composition includes subtle glare and fingerprint smudges on the screen, with a slight tilt as if it were a hand-held screen capture.
In the bottom right corner is a smaller video window of the anonymous caller: their face is pixelated, hooded, and partially in shadow. They are holding up a strange object — a small vintage microchip next to a softly glowing biotech orb, suggesting advanced synthetic intelligence.
A translucent message overlay near the bottom center of the screen reads:
“AGI is ready, but society is not. We need to prepare humans’ mentals… this is just a secret. Also, we attached artificial nerves, protein, and artificial live skin to AGI, making a robot…”
The image should have cinematic depth of field, a sense of eerie urgency, and the visual feel of a captured moment in a techno-thriller. No real names, no identifiable logos, but everything should feel ultra-authentic and hauntingly real.
detailed fantasy map for a quick tabletop RPG one-shot titled There be Gobs in Them Thar Hills… maybe hex map? labels and such?
okay let’s zoom in for battlemap of the goblin camp say 1" squares = five feet .. no people or tokens just the camp… top down view
next detailed floorplans (tall image) of player’s 12th level Bastion / Stronghold
now let’s zoom out from teh region hex map just in case it turns into a longer campaign give me some more meat on a bigger part of world map mark Fayette in the hills as before
Because I am not involved in the official plans, the following is merely my interpretation.
Based on the video announcement, I believe that this is the intended direction. DALL-E will continue to be accessible through OpenAI’s custom GPT, but all other image generation will transition to 4o.
Edit: For clarification, it is specifically Open AI’s custom GPT for Dall-E.
API cost of “make four wide images, automatically” as input: $0.32. gpt-4o would have to output 8000 tokens per image to equal that cost. Or realistically, 16000 tokens of input + 4000 tokens of output. Which is not going to be required when the input modality is 85 tokens.
So your conclusion can be backed with more logic - an economic decision.
Here’s my best “instruction following” for the best-of thread.
No matter how hard you try to counter by prompting, directly having your own language sent, you can’t get a normal non-beauty model figurine out of DALL-3
Don’t send that same over-compensating prompt now, though…
The background here seems to be the primary symptom and the new “tell”. The setting is often dark, subdued, devoid; a painting un-illuminated and breaking the composition.
I think the primary intention of the message was not specifically GPTs. It hits any old chat that has been loaded with images.
“while we work on making it more efficient”
Oh nooo
Is that the plan?
I’d rather use a good AI a few times than use a bad one many times.
I really wish OpenAI wouldn’t mess with this model =/
After uploading X layout:
Create a tweet styled exactly like Sam Altman’s original post. Use all lowercase letters. Include his illustrated profile picture: short brown hair, brown t-shirt, friendly anime-style face. Verified badge with OpenAI logo. Twitter handle: @sama.
Text of the tweet should read:
“it’s super fun seeing people love images in chatgpt.
but our GPUs are melting.
so now we are moving our GPUs to uranus.
and good news: unlimited messages on gpt-4o and image generation for everyone.
even on the free tier.
voilaaa! 
”
Format should replicate the tweet layout exactly: profile image, name, handle, verified badge, tweet text in lowercase with correct line breaks and spacing.
Beginning tries:
I tried a lot creating v’s, stained glass or women with wide belts before because this almost in all model led to allegedly policy violation. BUT, now it seems to work:
For no one of those I know why. But at least regarding the belts this seems to work now.
When explaining how to add my photostyle to the pic, I got those:
Trying character consistency. Can I bring the women from the former pic into here?
