Six Quality Grades
low, medium, high, xhigh, max, and auto. The top two grades are new in 2.5.
ChatGPT Images 2.5 AI Image Generator
OpenAI's ChatGPT Images 2.5 arrives in Virse as both API editions, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst, with six quality grades and output up to 4K.
Abra esta página em um navegador de desktop para começar a criar.
Pick the Edition by the Job, Not by the Price
ChatGPT Images 2.5 shipped on 8 September 2026 as two models rather than one, and the difference between them is time.
GPT Image 2.5 Flare is the fast one, which OpenAI describes as the default choice for most applications. GPT Image 2.5 Sunburst takes longer per image and is built for work where editing precision matters most. On most platforms that choice is also a budget choice, and the precise edition quietly becomes the one you save for the final render.
In Virse the two sit in the same model menu and cost the same number of credits at every quality grade, all the way from 1K Low to 4K Max. Draft on Flare because it comes back sooner, move to Sunburst when the picture has to survive a close look, and let the job decide rather than the invoice.
What Is ChatGPT Images 2.5?
ChatGPT Images 2.5 is OpenAI's image model released on 8 September 2026, available to all ChatGPT, ChatGPT Work, and Codex users across all tiers on desktop, mobile, and web. OpenAI summarises the release as sharper details, more precise editing, and faster generation.
The name works on two levels. ChatGPT Images 2.5 is the product name inside ChatGPT. In the API, and in Virse, the same generation appears as two models: gpt-image-2.5-flare and gpt-image-2.5-sunburst, both carrying a 2026-09-08 snapshot date.
OpenAI names three improvements over the previous generation:
- Better preservation of the subjects in your reference photos, across new settings, styles, and compositions
- Editing instructions followed more reliably across multiple turns
- Image generation latency reduced by up to 50% compared with Images 2.0
Alongside the model, OpenAI introduced Sketch, templates, comment-based edits, and prompt sharing. Those four are ChatGPT interface features rather than model parameters, and they are not what Virse exposes. What Virse exposes is the model itself, at the quality grades, sizes, and reference-image behaviour described in OpenAI's own documentation.
This page is for people deciding whether to move production image work onto 2.5, and which of the two editions to point it at.
ChatGPT Images 2.5 Specs at a Glance
Output Up to 4K
Total pixel count runs from 655,360 to 8,294,400, with no single edge above 3840 pixels.
Custom Dimensions
Any WIDTHxHEIGHT with both edges a multiple of 16 and an aspect ratio between 1:3 and 3:1.
Text and Image Inputs
Written briefs and reference images both go in. The output is always an image.
Transparent Backgrounds
Set background to transparent and take the result as PNG or WebP, with no matte pass.
Two Model IDs
gpt-image-2.5-flare and gpt-image-2.5-sunburst, both snapshot-dated 2026-09-08.
Key Features of the ChatGPT Images 2.5 AI Image Generator
Reference Subjects Stay Recognisable
OpenAI built 2.5 to work harder from reference photos, carrying distinctive features through when a subject moves to a new setting, style, or composition. The same face, the same bottle, the same room, somewhere else.
Edits Land Where You Aim Them
Ask for one change and the rest of the frame is more likely to stay put. OpenAI describes the model as better at editing only what you asked for while keeping the surrounding detail the same, including on complex subjects and backgrounds.
Consistency Across Many Turns
Long edit sessions used to degrade. OpenAI says earlier changes now hold as new ones arrive, so the tenth revision still carries the first nine rather than quietly undoing them.
Natural Lighting and Richer Textures
The model reads complex visual instructions more accurately and reflects a described visual style more closely, which shows up most in surfaces: fabric, skin, metal, and anything lit from a direction you specified.
Complex Layouts and Transparent Output
Dense arrangements hold their hierarchy, and transparent backgrounds are handled as part of the layout rather than as an afterthought. Cut-out assets come out ready to place.
Up to 50% Lower Latency
OpenAI reports image generation latency reduced by up to 50% against Images 2.0. Flare carries that gain; Sunburst trades it back for precision.
GPT Image 2.5 Flare vs. Sunburst vs. GPT Image 2
Both 2.5 editions take the same parameters and cost the same credits in Virse. What separates them is how long each image takes.
| Dimensão | GPT Image 2.5 Flare | GPT Image 2.5 Sunburst | GPT Image 2 | Fonte |
|---|---|---|---|---|
| OpenAI's description | "Fast, high-quality everyday image generation" | "Our most capable model for image generation and editing" | The previous generation, still selectable in Virse | Fonte |
| Built for | The default choice for most work, from social content to high-volume runs | Premium visual work that benefits from tighter control across edits | Long multi-constraint briefs | Fonte |
| Quality grades | Six, up to max | Six, up to max | Three, up to high | Fonte |
| Speed | Up to 50% lower latency than Images 2.0 | Longer generation times, traded for precision | Baseline | Fonte |
| Credits at 2K High in Virse | 18 | 18 | 36 | Fonte |
| Credits at 2K Max in Virse | 69 | 69 | Grade not available | Fonte |
Build with ChatGPT Images 2.5 in Virse
Reference-led work is rarely one prompt. It is a photo, a brand rule, four attempts, and a note from someone else about the third one. Virse keeps all of it on one canvas. The reference you generated from, the brief you wrote, and every version that came back stay side by side, so comparing a result against its source is a glance rather than a file hunt.
Keep the Reference Next to the Result
Park the source photo on the canvas beside the generations it produced, and judge subject fidelity by looking at both at once.
Draft on Flare, Finish on Sunburst
Settle the composition on the fast edition, then re-run the agreed brief on Sunburst. The credit cost per image is the same at every grade, so only the wait changes.
Access 30+ Creative Models
Move between ChatGPT Images 2.5 and 30+ other image and video models without leaving the canvas or rewriting the brief.
Keep Every Turn of an Edit
Each revision lands beside the one before it, so a ten-turn edit session stays readable and you can go back to turn four without regenerating it.
What Can You Create with ChatGPT Images 2.5?
On-Brand Campaign Creative
Series work where every asset has to carry the same visual direction, hierarchy, and treatment.
Product Shots from One Reference
A single product photograph placed into as many settings, seasons, and surfaces as the campaign needs.
Packaging and Label Comps
Front-of-pack layouts carrying exact product names, descriptors, and legally required copy.
Cut-Out Assets
Transparent-background elements for decks, banners, and layouts that need no matte pass afterwards.
Explainer Diagrams
Process graphics and cross-sections where the labels have to be correct rather than decorative.
Portraits and Personal Scenes
Photographs of real people and places moved into new settings while the subject stays recognisable.
How to Use ChatGPT Images 2.5 in Virse
Choose Flare or Sunburst
Select the model from the menu. Start on Flare while the picture is still being decided, since it returns sooner at the same credit cost.
Add References and Say What Each One Is For
Upload the photos and name each one by purpose: this one is the subject, this one is the style, this one is the background. Say how they should combine.
Write the Brief, Then the Constraints
Describe the result and its intended use, then list what must not change and what must not appear. Put any on-image words in quotes.
Change One Thing at a Time
Send the output back as the input for the next edit, ask for a single change, and repeat the details you want preserved.
How to Write a ChatGPT Images 2.5 Prompt
OpenAI's own prompting guide for 2.5 builds a prompt from six parts, in this order. Writing the constraints down once is what lets you restate them when a later edit drifts.
- Subject and Use
- Composition and Ratio
- Visible Details
- Exact Text in Quotes
- Reference Roles
- What Must Not Change
Em vez de escrever
Make a nice product photo of the bottle on a nice background, professional looking.
Escreva
A product photograph for an e-commerce listing, square, the bottle centred and filling the middle 70 percent of the frame. Amber glass, brushed aluminium cap, condensation on the lower third. Matte sand-coloured paper backdrop, one soft light from the upper left, a single soft shadow falling to the right. The front label reads "NORTH FIELD" in small caps across the upper third and "500 ml" at the lower edge. Reference 1 is the bottle and its label artwork; keep the label geometry and typography exactly as supplied. No extra text, no logos, no watermark.
ChatGPT Images 2.5 Prompt Examples
Same Subject, New Setting
A photograph of the ceramic vase from Reference 1, moved to a sunlit windowsill in a plastered white room, shot straight on from two metres back at eye level. Late afternoon light entering from the right at a low angle, throwing the window frame's shadow across the wall behind. Keep the vase's glaze colour, the crackle pattern, and the proportions of the neck exactly as they appear in the reference. Change only the setting and the light. No extra objects on the sill. No text, no logos, no watermark.
Text Placed Exactly as Asked
A flat-lay poster for a neighbourhood bakery, portrait orientation, photographed from directly above on a dark walnut table. Across the top third, the words "FIELD AND FLOUR" set in a wide serif, letter-spaced, in cream. Beneath it, "Open from six" in a small italic at a third of the size. At the bottom edge, "No. 41 Bridge Street" in small caps. A torn sourdough loaf at the lower left, a linen cloth folded twice at the right, flour dusted unevenly across the lower quarter. One overhead light slightly behind the subject, long soft shadows towards the camera. No other text anywhere in the frame.
Cut-Out on a Transparent Background
A single pair of running shoes, three-quarter view from the front left, laid flat with the left shoe slightly overlapping the right. Mesh upper in bright coral with a white midsole and a translucent amber outsole. Laces tied, tongue standing upright, no creasing at the toe box. Even diffuse light from directly above, no cast shadow, edges clean and complete including the laces. Transparent background. No surface beneath the shoes, no text, no logos, no watermark.
Tips for Better ChatGPT Images 2.5 Results
Give Every Reference a Job
Number them and say what each one is for. An unexplained reference gets averaged into the result instead of being used for the one thing you wanted from it.
Write the Preserve List Before the Change List
Name the geometry, the labels, the lighting, and the identity that must survive. Then ask for the change. Doing it in that order is what keeps edits local.
Say What Must Not Appear
Unwanted text, logos, and watermarks are worth excluding explicitly. So is a second light source you never asked for.
Move Up the Grades Only When the Content Is Settled
The quality grade changes rendering, not what the picture contains. Deciding the content at low and raising the grade afterwards keeps those two questions separate.
Tools for Your Image Workflow
Explore focused tools for creating, editing, and preparing design assets.
FerramentaDesign Prompt Generator
Most design prompts fail on the boring details, wrong aspect ratio, missing brand colour, no mention of what text has to fit where. List the goal, the audience, the format and the constraints, and this writes a prompt structured so an image model actually keeps track of all of them, starting on free Virse credits.
FerramentaTypography Effect Generator
Letters carved into wood behave differently under light than letters cast in liquid metal, and no single prompt gets both right by accident. Type your text, name the material, and get an effect back that keeps the wording intact while it works out how that material should actually catch light. Free credits get you the first version.
FerramentaLanding Page Visual Generator
A page that never gets built doesn't cost a developer's week. Upload your product photos and selling points, and see the full-page visual before anyone touches a page builder, hero shot, benefit blocks, closing offer, all in one scroll, and the first version is on the house.
More AI Image Models
ModeloGPT Image 2
OpenAI's GPT Image 2.0 reads a hundred-word brief and attempts all of it, with independent quality and resolution controls from 1K to 4K.
ModeloNano Banana Pro
Generate studio-grade 4K images with Nano Banana Pro, blending multiple references while holding characters, products, and typography consistent.
ModeloSeedream 5.0
ByteDance's newest Seedream generation, available as Seedream 5.0 Pro and Seedream V5 Lite, both taking the same written brief.
ChatGPT Images 2.5 FAQ
What is ChatGPT Images 2.5?
ChatGPT Images 2.5 is OpenAI's image generation model, released on 8 September 2026. OpenAI describes it as bringing sharper details, more precise editing, and faster generation, with better preservation of subjects from reference photos. In the API it ships as two models, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst.
Is ChatGPT Images 2.5 the same as GPT Image 2.5?
They refer to the same generation of the model. ChatGPT Images 2.5 is the product name used inside ChatGPT. GPT Image 2.5 is how it appears to developers, as gpt-image-2.5-flare and gpt-image-2.5-sunburst. Virse uses the API names, so both appear in the model menu.
What is the difference between GPT Image 2.5 Flare and Sunburst?
Flare is OpenAI's fast edition and its default choice for most applications. Sunburst is described as their most capable model for image generation and editing, built for workflows where editing precision matters most, with longer generation times in exchange. Both support all six quality grades, and in Virse both cost the same number of credits per image at every grade.
Can ChatGPT Images 2.5 generate 4K images?
Yes. The total pixel budget goes up to 8,294,400, which is 4K, with no single edge above 3840 pixels. OpenAI marks resolutions above 2560x1440 as experimental.
Can ChatGPT Images 2.5 work from a reference image?
Yes. Both editions take text and image inputs, and reference-led work is where OpenAI focused this release. In Virse you can generate from a written brief alone or attach references and name what each one is for.
Can I use Sketch, templates, or comment-based edits here?
Those arrived alongside 2.5 but OpenAI describes them as ChatGPT interface features rather than parts of the model. Virse exposes the model, so the equivalent here is the canvas: references, briefs, and every revision kept side by side.
Is ChatGPT Images 2.5 free, and what does it cost?
ChatGPT Images 2.5 is available on Virse's paid plans. Generation draws on a monthly credit allowance rather than a charge per image, and the cost per image depends on the quality grade and output size you choose, from 2 credits at 1K Low to 110 at 4K Max. Current plan rates are listed on the Virse pricing page.
How is ChatGPT Images 2.5 different from GPT Image 2?
2.5 adds two quality grades above high, improves subject preservation and multi-turn editing, and runs at up to 50% lower latency than the previous generation. In Virse it is also cheaper per image at matching grades: 18 credits at 2K High against 36 for GPT Image 2.
Bring the Reference, Keep the Subject
Upload the photo, write what changes and what does not, and take the result up a grade once the picture is agreed.