Published: September 9, 2026
Image Model — generation & editing
UPDATE — September 9, 2026: GPT-Image-2.5 is now available in both Flare and Sunburst variants across text-to-image and editing. Editing is the headline change: edits stay scoped to what was asked, subjects stay recognizable across settings and styles, and quality holds across long editing sessions.
This page covers the full GPT-Image-2 family: the original GPT-Image-2 release and the new GPT-Image-2.5 release (Flare and Sunburst variants).
GPT-Image-2 is OpenAI's flagship image model, strongest on photorealism and on text that stays legible — signs, UI, packaging, infographics.
The 2.5 release splits it into two variants: Flare for everyday speed, Sunburst for precision work, both covering generation and editing. The original GPT-Image-2 remains available and is still fine for straight generation, but 2.5 supersedes it for edit-heavy work.
Pick this family over Nano Banana or Flux when a job depends on rendered text or on repeated edits to the same subject.
Variants at a Glance
Variant |
Type |
Modalities |
Max Res. |
Talking Points |
|---|---|---|---|---|
| Variant | ||||
| GPT-Image-2 | Image | T2I, Editing | 4K | The original release — still fine for straight generation, superseded by 2.5 for edit-heavy work |
| 2.5 Flare | Image | T2I, Editing | 4K | Default for most work — creator and social content, product, visual search, prototyping, high volume |
| 2.5 Sunburst | Image | T2I, Editing | 4K | Slower, more precise across edits — campaign creative and polished product imagery |
Key Features
Image Editing
- Scoped Edits: Changes only what's asked; subject, composition and background stay untouched
- Subject Consistency: Distinctive features carry through across new settings, styles and compositions
- Long Sessions: Quality holds over successive turns instead of degrading
- Multi-Reference Input: Up to 16 reference images in one edit
Text-to-Image
- Instruction Following: Handles detailed visual instructions and real-world information more accurately
- Text Rendering: Near-perfect text on signs, UI and infographics — the model's standout strength
- Lighting & Texture: More natural lighting and richer texture than the original release
- Complex Layouts: Multi-element compositions hold, including transparent backgrounds
Technical Capabilities
Spec |
Details |
|---|---|
| Quality / Resolution | Six tiers (low → max); up to 4K, 3840px long edge |
| Aspect Ratios | 1:1, 3:4, 4:3, 9:16, 16:9, plus custom up to 3:1 |
| Input Modalities | Text-to-Image, Image Editing |
| Reference Images | Up to 16 per edit |
| Output Formats | PNG, JPEG, WebP — transparency on PNG and WebP |
| Variants | Flare (fast, default), Sunburst (precision, slower) |
Limitations
- Sunburst buys precision with generation time — wrong pick for high-volume or interactive work
- The top quality tiers push latency up sharply; use them for finals, not for iterating
- Aspect ratio caps at 3:1 — extreme panoramas and thin banners have to be composed wider and cropped
- Transparency needs PNG or WebP; JPEG flattens the background to solid
- No seed control, so a prompt won't reproduce the same image — lock a result by editing it forward
- GPT-Image-2 and 2.5 renders differ in lighting and texture; don't mix them inside one campaign
Prompting Tips
- Conversational full sentences beat keyword lists — the model reads a prompt the way a person would
- Put rendered copy in quotes and say where it sits: "the sign reads 'OPEN LATE' in bold condensed sans"
- For edits, name what should stay as well as what changes: "keep the background and lighting, replace only the label"
- Iterate on one image across turns rather than restarting from the prompt — 2.5 holds the subject steady
- Ask for a transparent background up front when the image is headed into a composite
- Explore on Flare, finish on Sunburst — same prompt, slow generation only on shipping frames