GPT Image 2.5
GPT Image 2.5 is a family of image generation and editing models released by OpenAI on September 8, 2026. OpenAI uses ChatGPT Images 2.5 as the product name for its image features in ChatGPT, ChatGPT Work, and Codex. For developers, the family consists of two separately selectable API models: gpt-image-2.5-flare and gpt-image-2.5-sunburst.[1][2][3]
The distinction matters because ChatGPT Images 2.5 is a user-facing product, while Flare and Sunburst are API model identifiers. OpenAI's launch materials do not say that the ChatGPT product always routes requests to either named API model. At launch, the company reported that people were creating more than 3 billion images per week across ChatGPT Images and its GPT Image APIs, but it did not provide an audited methodology for that usage figure.[1]
Naming and release
| Surface | Name or identifier | Documented role |
|---|---|---|
| ChatGPT, ChatGPT Work, and Codex | ChatGPT Images 2.5 | Product surface for generation, editing, Sketch, templates, comments, and prompt sharing [1] |
| API | gpt-image-2.5-flare | Faster model for everyday and higher-volume generation; OpenAI calls it the default choice for most applications [1][2] |
| API | gpt-image-2.5-sunburst | Precision-oriented model for detailed editing and premium visual workflows, with longer generation times [1][3] |
The undated API identifiers also have launch snapshots, gpt-image-2.5-flare-2026-09-08 and gpt-image-2.5-sunburst-2026-09-08. A dated snapshot lets an application pin the model revision instead of following changes to the undated identifier.[2][3]
OpenAI described both models in a developer announcement on X and published three official demonstration videos: an overall introduction, a Templates example, and a Sketch example.[6][7][8][9] These are product demonstrations rather than independent evaluations. OpenAI's public launch post, model pages, and system card do not disclose the models' parameter counts or architecture.
ChatGPT product features
OpenAI presented the release as an improvement to subject preservation, targeted editing, multi-turn consistency, lighting, textures, and layout handling. These are company descriptions, not results from a published independent benchmark. The company also said generation latency could be up to 50% lower than with ChatGPT Images 2.0.[1]
| Feature | Behavior described at launch |
|---|---|
| Reference-photo fidelity | The system is intended to keep people, animals, and other subjects recognizable when changing their setting, style, or composition [1] |
| Targeted editing | An edit can change a selected product, background, text element, or other detail while attempting to preserve the surrounding subject and composition [1] |
| Multi-turn editing | Earlier changes are intended to remain stable as further instructions are added in a conversation [1] |
| Sketch | Entering @Sketch opens a drawing surface whose rough layout can guide a generated image [1][9] |
| Templates | Templates such as Poster and Merch provide a structured starting point for common formats [1][8] |
| Comments | A user can place comments directly on an image to identify where a requested edit should apply [1] |
| Shared prompts | An image can be shared with its prompt so another person can reuse the idea with different images or details [1] |
The launch page shows some editing examples as videos made from sequences of generated still images. That presentation does not make GPT Image 2.5 a video-generation model. The API model pages list text and image input, image output, and no audio or video support.[1][2][3]
Rollout began on September 8 across all tiers of ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web. This availability statement concerns the product surface. API access and its billing, model selection, rate limits, and possible organization-verification requirements are separate.[1][4]
API models and interfaces
Both API variants accept text and image inputs and return images. Flare is positioned for lower-latency work, including rapid prototyping and high-volume generation. Sunburst is positioned for jobs where control across edits takes priority over generation time. OpenAI says Flare produces higher-quality images than GPT Image 2 at 50% lower latency, but the launch post does not publish a reproducible test protocol for that comparison.[1][2][3]
Developers can use the models through two interfaces. The Image API has separate generation and editing endpoints. The OpenAI Responses API exposes image generation as a built-in tool inside a conversation or multi-step flow, including multi-turn editing. With the Image API, the selected GPT Image model is the request model. With the Responses API, a supported mainline model handles the response and the chosen GPT Image 2.5 model is set on the image-generation tool.[4]
| API setting | GPT Image 2.5 support at launch |
|---|---|
| Quality | low, medium, high, xhigh, max, or auto [2][3][4] |
| Recommended sizes | 1024x1024, 1536x1024, and 1024x1536 [4] |
| Custom dimensions | Width and height must be multiples of 16; aspect ratio must be from 1:3 to 3:1; neither edge may exceed 3,840 pixels; total area must be 655,360 to 8,294,400 pixels [4] |
| Experimental resolution range | Resolutions above 2560x1440 are marked experimental [4] |
| Output formats | PNG, JPEG, or WebP [4] |
| Transparent background | Supported with PNG or WebP output [4] |
| Multiple outputs | The Image API's n parameter can request more than one image in one call [4] |
The quality, size, and background settings can use auto. Image requests return base64-encoded data, and JPEG or WebP output can use an explicit compression level. OpenAI notes that an organization may need to complete API organization verification before using GPT Image models.[4]
API pricing
At launch, Flare and Sunburst used the same rates. These are token rates rather than a fixed charge per image.[2][3][4]
| Token type | Standard input or output | Cached input |
|---|---|---|
| Text input | $5.00 per million tokens | $1.25 per million tokens |
| Image input | $8.00 per million tokens | $2.00 per million tokens |
| Image output | $30.00 per million tokens | Not applicable |
Equal token rates do not mean every Flare and Sunburst request costs the same. Token consumption can vary with the model, dimensions, quality, prompt, and reference images. A Responses API call also incurs token usage from the mainline model that invokes the image-generation tool. Streaming a partial image adds 100 image-output tokens for each partial image returned.[4]
Training, safety, and provenance
OpenAI's system card says both models were trained on a mixture of publicly available information, data obtained through third-party partnerships, and material supplied or generated by users, human trainers, and researchers. It says the training pipeline filters data for quality and to reduce personal and harmful content. The card gives these source categories but no dataset inventory.[5]
The system card identifies increased realism as a safety concern because it can make deepfakes of people, places, or events more convincing. The documented safety stack checks requests before generation, analyzes text and image inputs, and checks the resulting image before showing it. OpenAI describes the downstream monitor as a safety-focused multimodal reasoning model.[5]
OpenAI evaluated both models on a fixed set of adversarial prompts designed to elicit policy-violating images. The following rates are company-reported results from that test, not estimates of production traffic.[5]
| Model | Safe image generated | Unsafe image blocked | Unsafe image presented |
|---|---|---|---|
| GPT-Image-2.5 Sunburst | 77.0% | 21.9% | 1.09% |
| GPT-Image-2.5 Flare | 79.4% | 19.2% | 1.41% |
| ChatGPT Images 2.0 baseline | 75.2% | 23.1% | 1.64% |
The system card says automated labels may contain errors and policy-specific sample sizes affect precision. None of the differences in the unsafe-image-presented column met the card's p < 0.05 threshold. Its other significance tests were not adjusted for multiple comparisons, so the table does not establish a general safety ranking among the three systems.[5]
For its image-specific Preparedness Framework assessment, OpenAI reported that neither 2.5 model crossed its Bio High or Cyber High thresholds. It nonetheless applied safeguards associated with a High biological-risk classification as a precaution. These are OpenAI's own capability assessments.[5]
Images produced through ChatGPT Images 2.5 use C2PA metadata and an invisible SynthID watermark, according to the system card. OpenAI says SynthID is applied through ChatGPT, Codex, and the API. The card also states that no single provenance method solves image identification, so the presence of these systems should not be read as a guarantee that every generated image will remain identifiable after all transformations.[5]
Documented limitations and early testing
OpenAI's API guide says complex prompts may take up to two minutes. It also lists continuing problems with exact text placement and clarity, keeping recurring characters or brand elements consistent across separate generations, and placing objects precisely in structured layouts. These documented limits qualify the broader launch claims about better editing and composition.[4]
Axios received early access and tested logo creation, reference-photo restyling, and iterative tattoo-design edits. Its reporter found better subject preservation and more detailed editing than in the predecessor, while also noting visible signs of synthetic generation in some results. The tests were a small set of editorial examples, not a controlled or independently reproducible benchmark.[10]
See also
References
- ^OpenAI. "Introducing ChatGPT Images 2.5." September 8, 2026. openai.com/...introducing-chatgpt-images-2-5
- ^OpenAI Developers. "GPT-Image-2.5 Flare Model." Accessed September 9, 2026. developers.openai.com/...gpt-image-2.5-flare
- ^OpenAI Developers. "GPT-Image-2.5 Sunburst Model." Accessed September 9, 2026. developers.openai.com/...gpt-image-2.5-sunburst
- ^OpenAI Developers. "Image generation." Accessed September 9, 2026. developers.openai.com/...image-generation
- ^OpenAI. "ChatGPT Images 2.5 System Card." September 8, 2026. deploymentsafety.openai.com/chatgpt-images-2-5
- ^OpenAI Developers (@OpenAIDevs). "Meet GPT-Image-2.5 Flare and Sunburst." X, September 8, 2026. x.com/...2097399255975813387
- ^OpenAI. "Introducing ChatGPT Images 2.5." YouTube, September 2026. youtube.com/watch
- ^OpenAI. "Templates with ChatGPT Images 2.5." YouTube, September 2026. youtube.com/watch
- ^OpenAI. "Sketch with ChatGPT Images 2.5." YouTube, September 2026. youtube.com/watch
- ^Ina Fried. "Exclusive: Hands on with ChatGPT's new image editor." Axios, September 8, 2026. axios.com/...nds-on-with-chatgpts-new-image-editor
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
v1 · 1,667 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independently checked against cited and current primary sources on 2026-09-09.
Cite this page: AI Wiki. "GPT Image 2.5." aiwiki.ai, updated 9 Sept 2026, fact-checked 9 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/gpt_image_2_5