GPT Image 2.5 Arrives, Capable of Generating Fake GPT-6 Images

A new version of OpenAI’s image generation model appears to have emerged. According to reporting by QbitAI’s Wen Le, some users have received notifications about the new version inside ChatGPT. The leak comes at a stage when no official announcement has been confirmed. The name “GPT Image 2.1” had previously been expected. However, the actual label jumped straight to “2.5”. The jump in numbering suggests the scale of the changes is significant. The examples circulating so far are highly polished. Images purporting to recreate a GPT-6 launch event and executives’ communications are spreading. Notably, text reproducibility has advanced significantly from the previous generation.

How the Unannounced Model Was Revealed

Through Notifications

Reports say a new notification has appeared inside ChatGPT. It is not uncommon for names to be exposed without prior notice to users. This case is believed to have spread through a similar route. According to the QbitAI report, the wording of the notification itself has not been detailed. Even so, the emphasis is on the version number being “2.5”. That is beyond the predicted “2.1”. Some see this as including feature enhancements rather than minor fixes. Related moves have also been observed on external evaluation platforms. On LMArena, an anonymous model “luna-lisa-alpha” is drawing attention. There is speculation that this model is the new GPT Image. Its characteristics have been through a period of trial use. There are reports that generation speed is very fast. It is also rated as having more natural photographic texture. Face consistency is said to have improved when generating multiple images of the same person. Less garbled text inside images is also cited. The noise problem pointed out in GPT Image 2 is said to have been improved.

Why Text Rendering and Realism Have Improved

Significantly

What stands out in the examples is the accuracy of embedded text. Images imitating an official GPT-6 launch event are spreading. Presenters are depicted as if they were real people. Images imitating an all-hands message from Mr. Altman are also circulating. There are also recreated images of posts supposedly made by him. Apart from the portrait portions, no garbled text is visible. This level is approaching a stage where it could be misused for forgery. Examples of handwriting reproduction have also been shared. The test prompt specifies conditions in detail.

A photo of handwritten text in a realistic style, taken with a smartphone. The text is written in pencil on 8.5 x 11 inch ruled paper, the handwriting looks like Einstein’s handwriting, and the content is about the history of world currency. Give the handwriting clear Einstein-like variations. Add a slight coffee stain in the upper right corner.

The prompt specifies everything from paper size to writing instrument. It also includes handwriting characteristics and the location of stains as conditions. Comparisons with photos of actual handwriting have also been made. The paper texture and shading are close to a photograph. Pencil shading and interference with ruled lines are handled naturally. With previous generative models, long handwritten passages tended to break down. In these examples, line distortion and missing characters are suppressed. More stable landscape rendering has also been reported. Trees, houses, and animals are said to be less likely to break down in shape. Depth-of-field rendering is also said to have gained depth. In direct comparison with GPT Image 2, the difference is said to be clear. Outline distortion and unnatural textures have decreased.

Faster Speed and More Stable Depictions of

the Same Person Confirmed

Speed and stability matter greatly in practical use. Faster generation allows more iterations. This directly affects image selection for documents and mass production of ad creatives. Reproducibility of the same person matters for narrative generation. Uses that feature the same character across multiple scenes will expand. If changes in facial features are suppressed, retouching takes less effort. Stable text is effective for generating charts and mock screens. It becomes easier to imitate presentation materials and reproduce screen photos. Reduced noise improves reliability for photographic use. Grainy artifacts and color breakdown are said to be suppressed. If you are interested in prompt structure, comparison with the cases published in 500 GPT Image 2 Prompts Reverse-Engineered and Published on GitHub is helpful. Knowing trends in past prompts helps identify what has changed. It becomes easier to see under which conditions the new version performs best. Developers and designers will need to update their verification procedures. Text accuracy and person consistency should be added to evaluation criteria.

Why Image Generation Has Emerged as a Core Feature

There was a temporary settings change in the web version of ChatGPT. The image generation feature was reportedly moved up to the top of the main menu. It was reportedly placed next to new chats. It is now said to be hidden again. It may have been an experimental settings change. Even so, it is seen as showing where the operator is focusing. Some view image generation as becoming a key feature of next-generation models. There is also speculation that it will be positioned at the core of the GPT-6 era. Use that goes back and forth between dialogue and generation is expected to expand. Text editing and image revision can proceed on the same screen. The process of creating proposals and teaching materials could change significantly. The user base will also expand from specialists to the general public. Generation that requires no operational expertise will drive adoption. At the same time, distinguishing truth from falsehood will become even harder. Provenance management and watermarking technologies need to be developed in parallel.

Safety Challenges Posed by Easier Forgery

These examples show the difficulty of countering disinformation. Launch event images could spread if given context. Images imitating executive communications are also easy to trust. The more accurate the text, the harder it is to doubt. If portrait fidelity improves, visual identification becomes difficult. On social media, recompression and cropping occur during reposts. Traces of generation become even harder to see. Platforms need to strengthen detection technology and display-side measures. Invisible watermarks and provenance information for generated content are cited. Display design on the publishing side is also important. Mechanisms are needed to warn about images of unknown origin. Companies and news organizations are forced to review their verification procedures. Official announcements need systems that do not judge by images alone. If combined with video and audio generation, the impact will expand. Development of legal frameworks is also an issue. Handling of portrait rights and trademarks will be a prerequisite for business use. Balancing the scope of technology release with safety measures is the challenge.

Editorial Opinion

In the short term, we expect demand for fake-image detection and provenance management to grow. As text reproducibility in generated content improves, judging authenticity on social media by eye will become difficult. Platforms may rush to implement watermarking and provenance attribution. News organizations and corporate communications will be forced to update their verification procedures.

In the long term, we assess that image generation will become established as a core function of conversational AI. Document and ad production workflows will be reorganized around generation. Creative environments that require no operational expertise will drive adoption. Handling of copyright and portrait rights will become a condition for commercialization.

As a question from the editorial team, where does responsibility for authenticity lie. With the model provider, the user who published it, or the platform that distributed it. The focus will be on whether technological advances or institutional design catches up first. We believe the challenge is how to design a balance between convenience and safety.

References

• “新版GPT Image 2.5已经能伪造GPT-6发布会了”, by 闻乐 — 量子位, 2026-09-03T21:57:08.000Z (ARR)
• Source URL: https://www.qbitai.com/2026/09/483948.html