On September 8, 2026, Beijing time, OpenAI released ChatGPT Images 2.5, and simultaneously launched GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst in the application program interface. The new version is being rolled out to all levels of ChatGPT, ChatGPT Work and Codex users on desktop, mobile and web. OpenAI summarizes the improvements as clearer details, more accurate local editing, and faster generation speeds, but these performance judgments come from company announcements, and the release materials do not list a country-by-country availability schedule.
Reference image fidelity and creation entrance
OpenAI said that Images 2.5 can better maintain the identifying characteristics of people or objects when processing reference photos, and it is more likely to only change the part specified by the command, while retaining the rest of the subject, composition and brand processing. The company also says that early revisions across multiple rounds of dialogue are more likely to be consistent in subsequent edits, and image quality doesn't degrade to the same extent with each round. The official page illustrates these changes with examples, but does not disclose the complete test set, scoring method or item-by-item comparison of different tasks. Therefore, the report cannot equate product display with external evaluation results.
Four new creation portals have been added to ChatGPT. Sketch allows users to draw sketches directly, and then use line drawings and text descriptions together as a reference for generation; templates are oriented to common formats such as posters and product photos; picture comments are used to specify more focused modifications; when sharing pictures, users can choose to disclose the original prompts at the same time, allowing others to replace their own photos and details. The OpenAI release notes stated that Sketch is located in the @ menu of the message box on the mobile terminal, and image commenting and editing can also be entered from the full-screen image on the mobile terminal; the loading animation will vary depending on the platform and release batch, and the Snake mini-game is not available in Work mode.
API dual models and speed proposition
Developers get two new graphics models. OpenAI positions Flare as the default choice for most applications, saying it improves quality and editing capabilities while reducing wait times, and is suitable for social content, product experience, visual search, prototypes and high-frequency generation. Sunburst is aimed at creative and editing processes that require more detailed control, including creatives and product images, but officials also stated that it takes longer to generate. The positioning of the two models constitutes a product trade-off between speed and fine control, and the announcement does not provide a unified conclusion that applies to all workloads.
OpenAI claims that users generate more than 3 billion images every week through ChatGPT Images and the GPT-Image model in the application programming interface; the company also claims that Images 2.5 reduces image generation latency by up to 50% compared to Images 2.0. The former number does not come with statistical calibers, deduplication methods or third-party audits, and the latter number does not disclose the complete test configuration, request types and latency distribution. Therefore, what can be confirmed is the scale and decline limit disclosed by the company, and it cannot be inferred that every task, every region or every terminal will obtain the same speed change.
Token Prices and Security Stacks
The OpenAI pricing page lists Flare and Sunburst under the same standard rates: $8 per million image input tokens, $2 for cached image input tokens, and $30 for image output tokens; $5 for text input tokens and $1.25 for cached text input tokens. The units in the table are per 1 million tokens and are not a fixed selling price for a single image. The final cost of a request also depends on the actual input, output, size, quality and calling method; the release material does not provide a picture-by-picture bill comparison of the two models in common sizes, nor does it change the ChatGPT subscription price to the interface price.
The system card states that Sunburst uses multi-layered image security measures: requests are subject to policy checks before entering the generated model, input images and prompt words are jointly analyzed, and the generated images are checked again before display. The system card does not give an equally detailed layer-by-layer explanation of the security stack for Flare. The training data categories for the two models include public Internet information, information obtained by OpenAI through partners, and information provided or generated by users, manual trainers, and researchers; the company said that the processing process will filter personal information and use security classifiers, but has not disclosed an itemized list of data sets.
Assessment limits and subsequent verification
In fixed adversarial automated evaluations, system cards listed final unsafe output rendering rates of 1.09% for Sunburst, 1.41% for Flare, and 1.64% for the Images 2.0 baseline. The system card clearly states that the difference in this column does not meet the statistical significance threshold used; automatic labeling may have errors, and the sample size of each policy category will also affect accuracy. OpenAI also admitted that higher realism may produce more convincing political, sexual or other sensitive deepfakes in the absence of protection, and said that it will continue to use C2PA metadata and SynthID invisible watermarks, but there is no single solution for source identification.
The public page title and summary published by Axios on September 8 confirmed the launch of OpenAI's new image editor and said that its reporters conducted a hands-on task; the available public summary did not provide complete test data that could replace independent benchmarks. The next nodes that can be checked include the actual reach of each platform and region, whether Sketch and review entrances appear simultaneously, the real costs for developers at different sizes and qualities, and whether external evaluations can reproduce the fidelity, editing consistency and latency changes that OpenAI calls. Until such data emerges, performance, scale, and security conclusions must remain company-owned and system-card-qualified.