Continue GPT-6 ASTRA After the shockwaves it sent through the AI community, OpenAI may be preparing the next step in image generation. Recently, X account @chetaslua spotted a hidden promotional popup in ChatGPT’s Codex interface, with a title reading “Image creation got a major upgrade” and a “Try it” button, but the feature was locked behind a feature gate and unusable. The discovery quickly sparked discussion across the community, with observers dubbing the unreleased upgrade “ImageGen 2.5” or “GPT Image 2.5.”
Discovery process: the upgrade pop-up hidden inside Codex.
@chetaslua said he forcibly brought up a hidden onboarding popup in the ChatGPT Codex interface. The popup text reads: “Image creation got a major upgrade. Higher-quality results, faster generation, and smarter creative tools.” This is a typical product-upgrade intro, but it reveals no technical details at all: no model name, no benchmark data, no resolution, and no pricing or release date. Pre-placing but locking upgrade announcements like this in a product is not uncommon in the industry (Google and Apple have both been found multiple times with feature interfaces that weren’t yet enabled), but OpenAI preparing announcement material in advance within a developer-oriented product like Codex hints that this upgrade may be deeply integrated with Codex’s image generation capabilities.
🚨 Codex : ImageGen 2.5 announcement built in
forced the hidden modal on screen in ChatGPT
“Image creation got a major upgrade” “Higher-quality results, faster generation, and smarter creative tools”
gate is still off for everyone , hero image was uploaded Sep 1 pic.twitter.com/UWKfhNNs4i
— Chetaslua (@chetaslua) September 3, 2026
Two anonymous checkpoints: mona-lisa-1 and luna-lisa-alpha
Before the popup even appeared, traces of OpenAI’s testing had already surfaced on Image Arena (LMSYS image battle platform). Around August 9-10, an anonymous model named mona-lisa-1 began appearing in blind test battles.The tester discoveredIts output triggers OpenAI’s image verification tool, the model’s self-identification also reports as OpenAI, and the tokenizer matches the GPT family. However, its knowledge cutoff is around May 2025, which is older than the existing gpt-image-2 (cutoff December 2025), so it is interpreted as an early or side-line checkpoint.
Around August 19-20, a second anonymous model, luna-lisa-alpha, appeared. Testers reported that its performance was even better: a later knowledge cutoff, text rendering that was “outstanding,” output that was “hyper-realistic,” less graininess than mona-lisa-1, and more consistent character stability. Some testers said it was “better than GPT Image 2, but not to the point of being stunning.” The emergence pattern of these two checkpoints (anonymous Arena entry → OpenAI-style watermark → product-side leak weeks later) is consistent with the release cadence of gpt-image-2 back then.
Important reminderOpenAI has never acknowledged that these two checkpoints are associated with it, and neither has ever appeared on the official leaderboard. Whether there is a direct connection between mona-lisa-1 and luna-lisa-alpha and the possible upgrade behind the popup remains speculative at this point.
The name “ImageGen 2.5”: a community label, not an official designation
The names “ImageGen 2.5” or “GPT Image 2.5” are community-speculated labels, not official names used internally by OpenAI. The popup text itself makes no mention of any model name or version number. As of September 4, the only image model listed in OpenAI’s developer documentation, model lists, API endpoints, and changelog remains gpt-image-2 (snapshot gpt-image-2-2026-04-21), which launched on April 21, 2026 under the name ChatGPT Images 2.0. The model page URL for gpt-image-2.5 returns a 404. OpenAI’s GPT-6 Astra announcement, also released on September 3, makes no mention of any image model updates whatsoever.
Speculation on upgrade content: faster and smarter, but lacking concrete evidence.
The three promises in the popup (higher quality, faster generation, smarter tools) describe a user experience rather than a specific engine. Faster generation makes iterative editing more practical; “smarter tools” may point to product-level features such as region editing, aspect ratio control, and background processing. But there is no evidence to determine whether this is a base model upgrade, an improvement in inference efficiency, or a bundle of product-level features.
Looks like Codex is preparing a new image generation upgrade. I found strings for an imageGen25 announcement promising higher-quality results, faster generation, and smarter creative tools.
The interesting part is the internal “imageGen25” name. That could point to a new image… pic.twitter.com/2OprekON1w
— Token Gremlin (@TokenGremlin) September 3, 2026
The improvement in handwriting rendering is a noteworthy aspect of this upgrade. gpt-image-2 has achieved roughly 99% accuracy in rendering printed text, but handwriting—with its irregular glyph shapes, ink density, writing pressure, and paper texture—is a much harder domain. @chetaslua showcased a photorealistic output of handwritten text in an “Einstein style,” claiming the feature now handles handwriting fluidly. However, this is just a single demonstration from the person who discovered it, and it cannot be reproduced by others yet, so it should be treated as a directional signal rather than confirmation. If mature handwriting generation becomes a reality, it would greatly expand the application scenarios for AI image generation, from realistic documents and educational materials to personalized design.
🚨 GPT-Image-2.5 Output
I said it to write like einstein it did it without reference < i did not give image or anything >
left side is AI and right side is real photo from reddit to compare
Prompt : a photorealistic, taken by phone photo of a handwritten essay in… https://t.co/OAiHNaBruC pic.twitter.com/pP553WhCQa
— Chetaslua (@chetaslua) September 3, 2026
Conclusion: A leak doesn’t mean it will be released; next step is to watch DevDay.
As of September 5, OpenAI has not released any product named GPT Image 2.5 or ImageGen 2.5. All known evidence—two anonymous Arena checkpoints, a non-clickable popup, and a hero image—points toward OpenAI preparing some kind of image upgrade, but the specific content, name, and release timeline cannot be confirmed. Outsiders speculate that OpenAI may make an official announcement at the DevDay event in late September, but for now this is purely calendar-based conjecture.
In terms of actual impact on users, the current ChatGPT Images 2.0 (gpt-image-2) is already the top-tier image generation tool in the industry, ranking first in both the text-to-image and image editing leaderboards on Image Arena. On the API side, pricing ranges from $0.006 per 1024×1024 image (low quality) to $0.211 (high quality), plus input token fees, while on the ChatGPT side it’s covered under all paid plans. OpenAI also just phased out the older DALL-E 3 model on August 30, directing all users to ChatGPT Images. Given this product landscape, any subsequent upgrade would need to deliver sufficiently significant differentiation to be worth getting excited about.
This case again demonstrates the typical rhythm of AI product leaks: first, anonymous checkpoints appear on third-party testing platforms; then the hidden UI on the product side is uncovered; and only finally does the official announcement come. This time, we are currently between the second and third stages.
Source: KOCPC Chinese