GPT Image 2.5 is an advanced image generation and editing model from OpenAI, designed for sharper visual quality, stronger reference fidelity, and more precise refinement. It delivers more natural lighting and textures, preserves recognizable subjects more consistently, and follows complex visual instructions with improved accuracy across repeated edits.
Used to describe the generated images.
0/20000The 27:16, 16:27, 9:8 and 8:9 aspect ratios support 1K only. 2K and 4K are available for other aspect ratios.
Image resolution
Explore different use cases and parameter configurations
Used to describe the generated images.
0/20000Click to upload or drag and drop
Supported formats: JPEG, PNG, WEBP, JPG Maximum file size: 30MB; Maximum files: 16
Image for reference
The 27:16, 16:27, 9:8 and 8:9 aspect ratios support 1K only. 2K and 4K are available for other aspect ratios.
Image resolution
Explore different use cases and parameter configurations
Used to describe the generated images.
0/20000Click to upload or drag and drop
Supported formats: JPEG, PNG, WEBP, JPG Maximum file size: 30MB; Maximum files: 16
Image for reference
The 27:16, 16:27, 9:8 and 8:9 aspect ratios support 1K only. 2K and 4K are available for other aspect ratios.
Image resolution
Explore different use cases and parameter configurations
Used to describe the generated images.
0/20000The 27:16, 16:27, 9:8 and 8:9 aspect ratios support 1K only. 2K and 4K are available for other aspect ratios.
Image resolution
Explore different use cases and parameter configurations
Complete guide to using GPT Image 2.5
GPT Image 2.5 is OpenAI's image model, released in September 2026 in two versions: Flare, the smaller model, and Sunburst, the base model optimized for quality. On Kie.ai each version runs as two API models, one for text-to-image and one for editing with reference images.
Flare and Sunburst take exactly the same request, so switching between them is a change of model name. Text-to-image needs only a prompt; image-to-image also needs your reference images in input_urls.
| Parameter | Value |
|---|---|
| Models | Flare: gpt-image-2-5-flare-text-to-image, gpt-image-2-5-flare-image-to-image. Sunburst: gpt-image-2-5-sunburst-text-to-image, gpt-image-2-5-sunburst-image-to-image |
| Endpoint | POST /api/v1/jobs/createTask, then Get Task Details or a callBackUrl for the result |
| Prompt | Required; up to 20,000 characters |
| Reference images (input_urls) | Required for the image-to-image models; up to 16 image URLs |
| Aspect ratio | auto (default), 1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9, 27:16, 16:27, 9:8 or 8:9. 27:16, 16:27, 9:8 and 8:9 are available at 1K only |
| Resolution | 1K, 2K or 4K |
| Background | transparent, opaque or auto |
| Output in our tests on Kie.ai | PNG every time. 2736×1536 at 2K, 16:9; 3840×2160 at 4K, 16:9; 2368×1776 at 2K, 4:3; 1628×966 at 1K, 27:16; a transparent 1254×1254 PNG at 1K, 1:1 with background set to transparent |
| Pricing | Shown in the pricing card above |
OpenAI positions Flare as its small model, optimized for speed, with image quality comparable to GPT Image 2, and Sunburst as its base model, optimized for quality. On Kie.ai both take the same request. On LMArena, where both are still marked preliminary, Sunburst scores higher than Flare for text-to-image and single-image editing; the table shows the scores. In our own tests in October 2026 with the same prompts and settings, both followed a poster brief with correctly spelled text, kept a reference character through an edit, and returned transparent PNG stickers; Sunburst's poster showed more visible keys and tone-hole detail on the saxophone, and Flare split the title over two lines. Sunburst for quality; Flare when OpenAI's speed-oriented variant fits your use and GPT Image 2-level quality is enough — compare both on the same inputs. Pricing is shown in the pricing card above.
| Flare | Sunburst | |
|---|---|---|
| Models on Kie.ai | gpt-image-2-5-flare-text-to-image, gpt-image-2-5-flare-image-to-image | gpt-image-2-5-sunburst-text-to-image, gpt-image-2-5-sunburst-image-to-image |
| OpenAI's positioning | Small model, optimized for speed; image quality comparable to GPT Image 2 | Base model, optimized for quality; higher image quality than GPT Image 2 |
| Request parameters | Same as Sunburst | Same as Flare |
| Same poster prompt, 2K, 16:9 (our test) | Correct text; title on two lines | Correct text; title on one line; more visible keys and tone-hole detail on the saxophone |
| Same reference-image edit (our test) | Kept the reference robot's look | Kept the reference robot's look |
| Transparent background (our test) | Transparent PNG | Transparent PNG |
| LMArena text-to-image score (Oct 7, 2026, preliminary) | 1398±7 | 1425±7 |
| LMArena single-image edit score (Oct 6, 2026, preliminary) | 1481±5 | 1524±5 |
Each example below was generated with GPT Image 2.5 Sunburst on Kie.ai in October 2026 from the prompt described in its caption; images are shown as generated, resized for the web.
OpenAI's announcement, which calls the model Images 2.5, says it produces more natural lighting and richer textures. From a prompt describing late afternoon sun on a loaf of sourdough, Sunburst rendered the light falling across the table, the cracked crust, the linen weave and the wood grain. It also added a jug, a plate and a small bowl of salt that the prompt did not mention.
OpenAI says GPT Image 2.5 is better at preserving the subjects in your reference photos. Given our robot mascot as the reference image, Sunburst placed it on a kitchen counter next to a bowl of lemons and kept its colors, both antennas, its eyes and the lightning icon on its chest.
OpenAI says GPT Image 2.5 is better at editing only what you've asked for while keeping the rest the same. Asked to change the Lemon Tart price from 3.75 to 4.25, Sunburst changed that one price and left the title, the other dishes and prices, the drawing and the frame as they were (input on the left, edit on the right).
With background set to transparent, Sunburst returned a PNG with a transparent background, shown here on a checkerboard; the die-cut sticker kept its white outline.
OpenAI says developers build its Images API into products used by creative, marketing, retail and media teams. These are the kinds of work it describes for GPT Image 2.5.
Routine images for apps, posts and pages, the everyday image generation OpenAI describes Flare for.
Updating a single element, such as a product, background or piece of copy, while keeping the subject, composition and brand treatment around it.
Reference-led work where new versions of a character or product have to stay close to the original source image.
Edits where precision matters most, which is the kind of workflow OpenAI points Sunburst at.
The image-to-image models take reference images in input_urls, not image_urls. In our test, a request that used image_urls still returned an image but ignored the reference. Files that are not hosted yet can be uploaded first with the Kie.ai File Upload API (Base64, file stream or remote URL).
These four aspect ratios support 1K only. In our tests, choosing 2K or 4K with 27:16 failed with an "upstream service is temporarily overloaded" message, and the task records showed 0 credits consumed.
Set background to transparent when you need a cut-out; in our tests every result came back as a PNG, and the transparent ones had a real alpha channel. Kie.ai's playground notes that at 2K or 4K a transparent background also needs a prompt describing an isolated subject with no backdrop.
In our window-light test the model added a jug, a plate and a bowl of salt that the prompt did not mention, so review each image for extra objects before you publish.
All models in the Kie.ai market use the same authentication method: an Authorization: Bearer YOUR_API_KEY header with the key from your Kie.ai API Key Management page.
New Kie.ai users receive 80 free credits for testing. Your account has one credit balance that all generation services draw from; check it with the Get Remaining Credits API.
Kie.ai provides this model as an API with a playground; there is no app to download or install. There are no GPUs or model weights to manage.