ChatGPT Images 2.5 · Flare and Sunburst · released September 8, 2026
Posters, product shots, portraits, and UI mockups, generated and edited in the browser. Text renders cleanly, edits stay consistent, and every image is yours to use commercially. Below: what ChatGPT Images 2.5 changed, and what one image costs at each quality setting.
New accounts get free credits. No card required.
image-25.com provides AI image generation through third-party models. We are an independent product, not affiliated with OpenAI, Google, or any model provider.
Background footage: official ChatGPT Images 2.5 demo videos by OpenAI, shown to illustrate the model. Not produced by this site and not generated by our workbench.
No software to learn. If you can describe it, you can make it.
Subject, style, lighting, and any text you want in the frame.
Optional. Upload an image to guide the look or keep a character consistent.
Refine with follow-up prompts. Failed runs refund their credits automatically.
Model brief
OpenAI released ChatGPT Images 2.5 on September 8, 2026. It ships as two models with different jobs, and which one runs behind a tool changes what you get back.
ChatGPT Images 2.5 reached the API as two separate models. Flare is the one OpenAI names as the default; Sunburst trades generation time for tighter control across edits. Which one a tool runs is the single biggest factor in what you get back, so it is worth knowing both names.
Full comparison: GPT-Image-2.5 Flare vs Sunburst
OpenAI calls Flare the default choice for most applications: higher-quality images than GPT Image 2 at 50% lower latency. It is aimed at creator and social content, product experiences, visual search, rapid prototyping, and high-volume generation.
Sunburst is built for premium visual work that benefits from tighter control across edits, such as production-ready campaign creative and polished product imagery. It trades generation time for that extra precision.
Two of the before and after pairs OpenAI published with the release. The point of both is the same: the parts you did not ask to change do not change.
Before
After
Before
AfterOfficial ChatGPT Images 2.5 demo images by OpenAI, shown here to illustrate the model. Not produced by this site and not generated by our workbench.
Source: OpenAI, Introducing ChatGPT Images 2.5, and OpenAI API pricing documentation. Read September 9, 2026.
ChatGPT, GPT Image, and OpenAI are trademarks of OpenAI. This page describes their model for reference. We are an independent product and not affiliated with OpenAI.
Six things you can rely on, none of them a roadmap promise.
Headlines, signage, packaging copy, UI labels. Flagship models place readable type where you ask for it, in multiple languages, without the garbled letters that made earlier AI images unusable for real layouts.
Natural light, true color, believable skin and materials. Output reads as studio photography, not as an obvious render with the waxy AI cast.
Change one thing and only that thing changes. Faces, products, and layouts stay pinned across edits and variants, so you can iterate toward a final instead of rerolling from zero.
Upload an image to guide the look or keep a character consistent across generations. Describe the change; the rest of the frame stays put.
A menu of named image models, each priced in credits per image. Switch per generation. Stronger models and higher quality settings cost more credits, never a subscription.
A failed generation refunds its credits automatically. You only ever pay for images you actually receive.
Official ChatGPT Images 2.5 demo images published by OpenAI with the release. Four things worth looking at closely: lettering that stays legible, layouts that hold their structure, style range, and photoreal output.
Nine mid-century modern posters in a three-by-three grid, each pairing bold lettering with a geometric illustration. Nine separate blocks of type, all still readable.
Eight vintage-style national park stamps, each carrying its own park name in small type without the letters breaking down.
A vintage-style poster reading "ChatGPT Stickers" and "New sticker pack" in large yellow and white letters over a dark cat illustration.
Official ChatGPT Images 2.5 demo images by OpenAI, shown here to illustrate the model. Not produced by this site and not generated by our workbench. Your own results depend on the model and quality setting you pick.
Four jobs this workbench does daily.
Build banners and social posts that come out of the generator with a readable headline, supporting line, and button already in place. Ask for a clear layout and the model follows it: headline top, product center, call to action bottom. Need five variants for a test? Change only the headline text between runs and keep the rest of the composition identical, so you can ship a whole test matrix without opening a design tool.
Run this in the generator
Start from a clean studio render of your product, then place the same product in a kitchen, on a trail, or in a lifestyle scene. Shape, color, and branding stay stable while the world around it changes. When you need a dozen images for a product page and there is no budget for a shoot, this is the fastest honest way to get them.
Run this in the generator
Upload a photo and change the background, the lighting, or the whole setting while the face stays recognizably the same person. Turn a phone selfie into a studio headshot, a cinematic portrait, or something stranger. Because edits are targeted, you refine the same image step by step instead of starting over each time.
Run this in the generator
Describe a screen, get a screen: title at the top, content in the middle, button at the bottom, labels readable. It will not replace your design tool, but for pitching an idea or filling a deck before the real design exists, a generated mockup that respects structure beats a whiteboard photo every time.
Run this in the generatorFour habits that consistently improve results, from our prompting guide. →
Put text you want in the image inside quotation marks and say where it goes. Quoted strings render as literals, not suggestions.
Ask for a headline, or a headline plus one short line. More copy elements means more chances to slip; add the rest in an editing pass.
Bold condensed sans, thin elegant serif, hand-painted sign lettering. Style words steer the render; vague requests produce mush.
Small type is where models still fail. A headline across the top third renders far more reliably than a caption in a corner.
The generator runs on a menu of named image models, each shown with its credit price per image. When a stronger model ships, we add it to the menu and you switch when you want to. Your prompts, history, and credits carry over unchanged. Image 2.5 is our product name, not a model name; we are an independent workbench and every model in the menu is labeled with exactly what it is, including its API model id.
You buy a pack once and spend it at your own pace. One image costs a fixed number of credits shown next to the model you pick, a failed run refunds itself, and there is no renewal date to remember. Image generation is bursty work: a campaign week, then quiet. A subscription bills you for the quiet weeks; a pack does not. New accounts start with 15 free credits so you can judge the output before paying anything.
Monthly, yearly, or pay once. An image costs 4–84 credits depending on the model and size you pick, and failed generations refund automatically.
Light, steady image work every month
For creators who publish every week
High volume for teams and heavy users
Every new account starts with 15 free credits, no card required.
Everything about the models, the credits, and what you can do with the images.
Image 2.5 is an independent AI image workbench: write a prompt or upload a reference, generate and edit in the browser, pay with credits, and use every output commercially. Image 2.5 is our product name, not a model name. The model called ChatGPT Images 2.5 belongs to OpenAI, released September 8, 2026, and shipped in the API as gpt-image-2.5-flare and gpt-image-2.5-sunburst. We are not affiliated with OpenAI or any model vendor.
Sign in, type what you want, pick an aspect ratio and resolution, and hit generate. The model and its credit price sit on the prompt box before you run, and the result lands in the same panel. Everything happens in the browser: no API key, no install, nothing to configure.
The one you pick from the model menu. Each entry shows the model's full name and its credit price per image, so you always know what made your picture and what it cost. We name real models rather than saying advanced AI, and you can switch per generation — nothing is locked behind a plan tier.
Auto, plus fourteen fixed ratios: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5, 21:9, and the extreme banner shapes 4:1, 8:1, 1:4, 1:8. Resolution options come from the model you picked and sit next to the ratio — we only list what that model actually accepts.
Yes. Attach it as a reference (up to 4 MB) and describe the change — swap a background, restyle a product shot, correct one detail. Targeted edits keep the rest of the frame intact, which is what makes iterating practical instead of re-rolling from scratch.
Yes, and this is where the current generation improved most. Ask for exact wording and placement and it usually lands legibly, including non-Latin scripts. Very small type, dense paragraphs, and long strings are still the hardest cases, so check the output before you ship it.
It depends on the model and the size. The lightest model is 4 credits an image. GPT-Image-2.5 Flare, which runs at the medium quality tier, is 12 credits at 1K and 22 at 4K; Sunburst runs at the high tier and is 44 at 1K, 84 at 4K. The exact price sits next to each model in the menu and follows the size you pick, so there is never a surprise. A failed or timed-out generation refunds its credits automatically — you only pay for images you actually get.
Every new account starts with 15 free credits, enough for a full GPT Image 2.5 image, no card required. After that the smallest credit pack is $28.99 for 1,000 credits, or $11.99 a month for 500 credits if a subscription suits your rhythm better.
Monthly plan credits stay valid 30 days from each grant. Yearly plans hand you all twelve months upfront, valid for a year, so unused credits carry over month to month. One-time packs last 60 days for Sprint and a full year for the rest. Subscriptions are self-serve: cancel anytime under Settings → Billing and keep the credits already granted.
Every paid plan includes commercial use, and we add no visible watermark or logo to your downloads. Model providers may still embed invisible provenance signals — OpenAI, for one, attaches C2PA metadata to images made with its tools. You stay responsible for what you put in the frame: trademarks, real people's likenesses, and brand assets follow their own rules.
Yes. Every generation is tied to your account and appears only in your own history — nothing is posted to a public feed or gallery. Your prompt and any reference you upload go to the model provider solely to render that image, and you can ask us to delete your data at any time.
Make an account, spend the free credits, and judge the output against whatever you use today.
Product updates, guides, and engineering notes from the team.
OpenAI released ChatGPT Images 2.5 on 8 September 2026 with two API models. The quality labels moved: medium now costs a quarter of what it used to.
Image 2.5 TeamAn image costs 4 to 84 credits, by model and size. Three subscription tiers, four one-time packs, and every credit carries a stated expiry date. The parts most tools bury.
Image 2.5 TeamGarbled letters are a prompting problem as much as a model problem. Five habits that make headlines, signage, and UI labels come out readable.
Image 2.5 Team