Img ai · Create, refine, export
Turn a brief into a useful visual with GPT Image 2.5. Choose Flare or Sunburst, generate from text, or refine an image with references and a clear editing brief. Explore the models behind the Images 2.5 release in one workspace.
Image studio
GPT Image 2.5Choose Flare for quick iterations, or Sunburst when finer detail matters.
From brief to finished visual
Illustrative concept generated with Codex. Your output appears here; results vary.
Keep the parts that work.
Create a first version, choose Edit this image, then describe one change. Describe what to preserve and what to replace.
Sources checked September 9, 2026 · Official references
GPT Image 2.5 is OpenAI’s image model family launched on September 8, 2026, with two API variants: Flare and Sunburst. ChatGPT Images 2.5 is the related experience inside ChatGPT. Img ai provides an independent generation and editing tool for these models; it is not ChatGPT or an official OpenAI website.
The practical question is whether an output survives the next revision. OpenAI describes improvements to image fidelity, editing consistency and generation speed. For a product campaign, that means trying to preserve a bottle’s proportions while changing its setting; for a poster, it means checking that a revised background does not disturb the headline. These are useful evaluation criteria, not guarantees for every prompt. Read OpenAI’s release notes.
Use Flare to explore directions, make routine campaign variations and test a new prompt. OpenAI positions it as the default model for most applications. It supports the same generation and editing controls exposed in this workbench.
Consider Sunburst for intricate product details or a tightly specified revision. OpenAI positions it around premium visual work and extra fidelity, with longer generation times. Choose it because the detail matters to the deliverable, rather than simply because it is another model.
These original Codex-generated illustrations show the intended deliverables for the prompts below. They are not measured GPT Image 2.5 outputs or a head-to-head benchmark. Run the prompts in the workbench to evaluate your own results.
Plan the material, light and crop before generating. A believable shadow and an uncluttered surface often matter more than adding another style adjective. Use your own product reference when the shape must remain recognizable.
Create a premium skincare campaign photograph. One unbranded amber glass pump bottle on an ivory limestone plinth, warm morning side light, a soft olive branch shadow, realistic glass reflections. Landscape 4:3 composition with generous negative space on the left. No text, no logos, no extra bottles.
Put the headline in quotation marks and keep it short. Specify where the words belong, then inspect spelling, spacing and contrast at the final display size. For a reusable campaign, keep editable typography in your design file too.
Design a finished editorial coffee poster, landscape 4:3. Sculptural white ceramic espresso cup and saucer on a terracotta tabletop, dramatic morning shadow, cream background. Set exactly "SLOW MORNINGS" in large elegant dark brown type across the top. Only these two words; no other lettering, no logos, no watermark.
Start from a result you like, then request one change at a time. This makes it easier to spot unwanted changes to the product silhouette, camera angle or composition. Describe the region to change and explicitly ask to preserve the rest.
Keep the same amber pump bottle, limestone plinth, bottle proportions, framing and camera angle. Replace the warm ivory background with a muted slate blue studio background. Keep the product label area blank and preserve realistic glass reflections. Do not add objects or text.
Describe the subject, composition, light and output context. Put exact text in quotation marks. Start with one output while checking a new brief.
Choose Flare or Sunburst, an aspect ratio and 1K, 2K or 4K resolution. Start at 1K for an initial draft and use a higher resolution for a larger final image.
Switch to Edit and upload references. Say what must remain and what may change. Describe one change at a time and inspect the result before the next revision.
Check typography and product details. Use Edit this image for another revision or download the result. Keep one change per iteration when you need to understand what improved.
The workbench accepts up to 16 references and produces one image per request. Choose from 13 aspect ratios and 1K, 2K or 4K resolution. Auto lets the model choose the ratio. Local uploads are limited to 20 MB each. These controls follow what the model exposes; ChatGPT’s Sketch and comment interface are separate product features.
The controls and credit costs below are the ones Img ai exposes for each model. Capability claims about edit quality come from official positioning, not from an independent benchmark, so keep the brief, references, aspect ratio and resolution comparable when testing.
| Specification | GPT Image 2.0 | GPT Image 2.5 |
|---|---|---|
| Model variants | GPT Image 2.0: one model, used for text-to-image and image editing. | GPT Image 2.5: two variants, Flare and Sunburst, each with generation and editing. |
| Resolution | 1K, 2K and 4K. | 1K, 2K and 4K. |
| Aspect ratios | 16 ratios including Auto, with limits above 1K: Auto is 1K only, 1:1 is unavailable at 4K, and 5:4, 4:5, 3:1, 1:3 and 9:21 stay 1K only. | 13 ratios including Auto, all selectable at every resolution. |
| Reference images per request | Up to 8 references in the Img ai workbench. | Up to 16 references, and a finished result can become the next reference. |
| Credits per image on Img ai | 10 at 1K, 12 at 2K, 18 at 4K. | 20 at 1K, 30 at 2K, 40 at 4K, for either variant and either mode. |
| Prompt length | Up to 20,000 characters. | Up to 20,000 characters. |
| Editing and speed | The established baseline; judge preservation against your own reference, not a version number. | OpenAI reports better consistency through edits. Sunburst trades longer generation for precision; no fixed completion time is promised. |
For a first campaign direction, start with Flare. When a near-final output fails on small material details or a precise edit, test Sunburst on that same brief. There is no benefit in paying for four variations before checking whether one obeys your requirements. Both variants expose the same controls here; their intended precision and latency tradeoff is the reason to compare them.
Explore the GPT Image 2 workflow or compare all Img ai models. For another reference-driven image workflow, see Seedream 5.
The workbench shows your Img ai credit charge before submission. Each image costs 20 credits at 1K, 30 at 2K or 40 at 4K, for both Flare and Sunburst, generation and editing. Credits are reserved once per request; retrying a saved request reuses its identity. A confirmed failed task returns its credits. If a provider response is uncertain, the request stays saved for recovery rather than starting another paid generation.
Img ai credits pay for the complete workflow, including generation, editing and saved task recovery. Credit packages are separate from ChatGPT subscriptions and any direct model API billing. View Img ai plans and credits.
GPT Image 2.5 is the image model family introduced alongside ChatGPT Images 2.5 on September 8, 2026. Its API variants are Flare, intended for most applications, and Sunburst, intended for work that needs more precision and allows longer generation times.
They refer to the same release family, but to different surfaces: ChatGPT Images 2.5 is the ChatGPT product experience; GPT Image 2.5 Flare and Sunburst are API models. Img ai is an independent service that provides those models through its own workbench. A ChatGPT subscription does not pay for Img ai credits.
Start with Flare when you are exploring ideas or producing routine variations. Consider Sunburst when small details and carefully scoped edits justify extra waiting. Compare both with the same brief and settings before committing a campaign budget; model positioning is not a guarantee that every output will be better.
Yes. Choose Edit, upload one or more references and describe what to change and what to preserve. You can use a completed result as the next reference. Up to 16 references are supported; Img ai accepts PNG, JPEG and WebP files up to 20 MB each.
Choose 1K, 2K or 4K resolution and one of 13 aspect ratios, including Auto. The model determines the exact output dimensions. Img ai does not currently expose a separate transparency or export-format control.
This page offers prompt-based generation and editing, reference uploads and result-to-edit reuse. ChatGPT Sketch, shared templates and its comment editing interface belong to the ChatGPT product and are not reproduced here.
Generation and editing use Img ai credits. The workbench shows the credit charge for your selected settings before submission. Each request creates one image: 1K costs 20 credits, 2K costs 30 and 4K costs 40, for either variant and either mode. Review the Img ai pricing page for credit packages.
Inspect text, product details, faces, edges and any factual information. Use reference material you have permission to use and review the applicable service terms. This tool does not verify trademarks, ownership or the truth of claims printed inside an image.
Reviewed September 9, 2026. Model capabilities and availability can change; these are the primary references used for this page.
Start with one clear brief. Refine the details that matter.
Try GPT Image 2.5