OpenAI introduced its GPT Image 2.5 generation on September 8, 2026, bringing faster production, more precise editing, stronger subject preservation, and greater control over image quality. For developers, the release includes two API models rather than a single all-purpose option: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.
The distinction matters. Flare is optimized for speed and everyday image creation, while Sunburst prioritizes quality and editing precision. Both models accept text and image inputs, generate new visuals, edit existing images, and support transparent backgrounds.
Together, they represent a shift from treating AI image generation as a one-prompt novelty toward using it as a practical, iterative creative workflow. OpenAI confirmed the two models and their availability through the Images API and Responses API in its official API changelog.
What Is GPT Image 2.5?
“GPT Image 2.5” is best understood as the name of OpenAI’s latest image model family. In the API, developers choose between two specific model IDs:
gpt-image-2.5-flare
gpt-image-2.5-sunburst
Both can generate images from written descriptions and make natural-language edits to reference images. The models are designed to follow detailed visual instructions while preserving important elements such as a person’s appearance, a product’s shape, the composition of a scene, or an established artistic style.
According to OpenAI’s GPT Image 2.5 prompting guide, Flare is the smaller model and is optimized for speed, with image quality comparable to GPT Image 2. Sunburst is the quality-focused model and offers higher image quality than GPT Image 2. Both are intended to deliver more precise edits and better subject preservation.
Flare vs. Sunburst: What Is the Difference?
The two models share the same API features, but they are intended for different production needs.
Model | Main priority | Suitable use cases |
GPT-Image-2.5 Flare | Speed and efficient iteration | Social posts, content drafts, rapid prototyping, visual search, product features, and high-volume generation |
GPT-Image-2.5 Sunburst | Quality and precise editing | Campaign assets, polished product imagery, complex compositions, and premium creative work |
Flare is the more practical starting point when an application needs to generate many images or give users quick feedback. If an existing GPT Image 2 workflow already meets its quality requirements, OpenAI recommends testing Flare to see whether it can maintain acceptable quality while reducing latency.
Sunburst makes more sense when the image must meet stricter visual standards or when Flare cannot handle a complex edit reliably. OpenAI recommends evaluating Sunburst first for demanding workflows, then testing the same prompt with Flare to determine whether the faster model can still meet the required quality level.
Importantly, OpenAI does not claim that one choice will be faster or better for every request. Results depend on the prompt, reference images, output dimensions, and selected quality setting, so production teams should evaluate both models on representative workloads.
More Precise, Localized Editing
One of the most useful improvements in GPT Image 2.5 is its ability to understand what should not change.
Image models can introduce unwanted changes during editing. A request to replace a chair might also modify the room’s lighting. Removing an object could alter a nearby person’s face, pose, or clothing. Changing a product background might distort the packaging or label.
GPT Image 2.5 is designed to keep edits more localized. This makes it better suited to workflows in which visual consistency is essential. A retailer could isolate a product on a transparent background without redesigning the product itself. A marketing team could update a line of text while preserving the surrounding campaign layout. A designer could change the material of an object without rebuilding the entire scene.
The models can also work with multiple references. One image might provide the subject, another the setting, and others the clothing or style. For dependable results, the prompt should state clearly what each reference contributes.
Better Multi-Turn Image Workflows
GPT Image 2.5 supports iterative editing through the Responses API. Instead of trying to describe the final image perfectly in one prompt, users can create an initial version and refine it through a series of focused requests.
For example, a user could generate a product advertisement, change the weather in a follow-up request, translate the headline, and then adjust one object in the background. A conversational workflow makes it easier to refine an idea without restarting from scratch each time.
OpenAI’s image generation documentation recommends the Images API for straightforward, single-request generation or editing. The Responses API is better suited to conversational products and multi-step workflows because it can retain image inputs and outputs within the interaction context.
Flexible Quality, Size, and Background Settings
GPT Image 2.5 supports six quality settings: auto, low, medium, high, xhigh, and max. The xhigh and max settings extend the range available in GPT Image 2, although a higher setting does not guarantee a better result for every prompt.
Developers can also request custom image dimensions. Common supported sizes include:
1024×1024 for square images
1536×1024 for landscape images
1024×1536 for portrait images
2048×2048 for 2K square images
3840×2160 for 4K landscape images
2160×3840 for 4K portrait images
Custom dimensions must use multiples of 16. Neither edge can exceed 3,840 pixels, the aspect ratio must remain between 1:3 and 3:1, and the total image size must fall between 655,360 and 8,294,400 pixels. OpenAI currently classifies resolutions above 2560×1440 as experimental.
Both models support transparent backgrounds. To generate genuine alpha transparency, developers must set the background to transparent and choose PNG or WebP as the output format. JPEG does not support transparency.
GPT Image 2.5 API Pricing
Flare and Sunburst currently use the same per-token rates:
Text input: $5 per million tokens
Cached text input: $1.25 per million tokens
Image input: $8 per million tokens
Cached image input: $2 per million tokens
Image output: $30 per million tokens
These rates do not mean that both models will always cost the same per image. Token consumption can vary by model, output size, quality setting, prompt, and reference inputs. Responses API requests may also include charges for the main language model operating the image generation tool.
Because image costs depend on actual token usage, developers should check the response data and the current OpenAI API pricing page rather than relying on a fixed per-image estimate.
How to Get Better Results
Effective GPT Image 2.5 prompts should describe the intended image before listing the details. A strong prompt normally defines the subject, composition, style, lighting, required text, and any important constraints.
For an edit, state both what should change and what must remain untouched. “Replace only the white chairs with wooden chairs while preserving the lighting, shadows, camera angle, and all other objects” is more reliable than simply asking the model to “make the chairs wooden.”
It is also better to refine one element at a time. After every generation, check text accuracy, subject identity, product geometry, layout, and transparency before continuing.
Final Thoughts
GPT Image 2.5 is not simply an upgrade in visual sharpness. Its more significant contribution is greater control: creators can choose between speed and premium quality, preserve important details across edits, work iteratively, and generate assets in flexible sizes with transparent backgrounds.
Flare should suit most fast-moving and high-volume workflows, while Sunburst is the stronger candidate for demanding creative and editing tasks. Neither removes the need for human review, but both make it easier to move from an initial idea to a polished, usable visual without repeatedly rebuilding the image.