GPT Image 2.5 Makes You Pick a Model Before You Make a Picture

person holding white ipad with black case

On September 8, 2026, OpenAI stopped shipping one image model. ChatGPT Images 2.5 landed that day, and behind the friendly product name the company quietly split its image engine into two separate API models: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. One is built to be fast. The other is built to be precise. You now choose which tradeoff you want before you make a picture, and that small change matters more to working creators than any single benchmark OpenAI put on the slide.

There is a second change hiding under the same release, a Sketch tool that lets you draw the layout yourself. Together the two updates say something about where AI image tools are heading: away from “type a prompt, pray for a good roll,” and toward giving you the controls a designer would actually expect.

What actually shipped in 2.5

The consumer app is still called ChatGPT Images. In the chat window you will not see the word Flare or Sunburst at all. The split lives in the API, where developers and anyone wiring these models into a workflow now pick between the two by name, per OpenAI’s own breakdown.

Flare is the default. OpenAI describes it as its fastest model for high-quality, everyday image generation, and says it delivers images at two to four times the speed of the previous GPT Image 2, cutting latency by up to 50% compared with Images 2.0. Sunburst is the slower, more deliberate one, aimed at production-ready campaign creative and polished product imagery where editing precision matters most. Independent testing from CellCog clocked Sunburst at roughly 1.5 to 2 times longer than Flare across every quality setting.

The scale of this is easy to miss. OpenAI says people now generate more than 3 billion images a week on the platform, which is exactly why splitting the model in two is a business decision, not a vanity one. Most of those 3 billion are throwaway drafts. A tiny slice are finals someone will actually publish or print.

Flare or Sunburst: pick by the job, not the prompt

Here is the practical way to think about the two models. Stop asking which one is “better.” They are tuned for different moments in the same project.

Reach for Flare when you are exploring. Thumbnail concepts, social variations, mood boards, rapid what-if drafts, anything where you will generate twenty options and keep two. Speed is the whole point, and paying for precision you are about to throw away is waste.

Switch to Sunburst when you have locked the concept and you are editing toward a final. Its real advantage is multi-turn editing: OpenAI built it to change only what you asked while keeping everything else in the frame stable, so a fifth edit does not quietly redraw the face or shift the logo. That consistency across repeated edits is the thing that breaks in most image tools, and it is the difference between a usable product shot and a near-miss you have to redo. If you have used the editable-layer approach in models like Seedream 5.0 Pro and Reve 2.1, Sunburst is chasing the same goal from a different direction: keep the parts you like fixed while you change one thing.

The workflow that gets the most out of the split is a two-stage one. Draft wide and cheap on Flare, then move your one chosen composition to Sunburst and refine it there. Treating both as the same tool, which the chat interface tempts you to do, leaves the value on the table.

The Sketch tool is the quieter, bigger story

Type @Sketch in a ChatGPT message and you get a small canvas to draw on. Rough it out however you like: a box where the product goes, a circle for a face, a line marking the horizon, some scribbled placement for text. The model reads that drawing as a visual reference and builds the image around your layout instead of inventing its own. OpenAI ships starting templates for common jobs like posters and merch, so you are not staring at a blank square.

This is the part that changes how you work. Prompting with words alone means describing spatial relationships in sentences, which humans are bad at and which is why “put the logo top-left, keep the subject centered, leave room on the right for copy” so often comes back wrong. A ten-second sketch communicates all of that instantly. For anyone who has fought a text prompt to place elements where they belong, this is the control returning to your hand. It also lowers the barrier for creators who think visually but do not think in prompt syntax, which is most creators. The Figma design agent made a similar bet earlier this year, that the useful AI is the one that speaks the visual language you already work in rather than forcing everything through a text box.

Sketch does not make you a compositor. It gives the model a scaffold. But a scaffold is exactly what pure-prompt generation has always lacked.

What it costs, and why the split changes your bill

Both models carry the same token pricing, which trips people up: $8 per million input tokens and $30 per million output tokens, with cached input at $2. In practice CellCog puts a single 1024×1024 image anywhere from about $0.006 at low quality to roughly $0.211 at maximum quality. Same rates on both models. The difference in your monthly bill comes from how you use them, not from a price gap between them.

That is where a little operations discipline pays off. Running content pipelines for years taught me one durable rule about metered tools: match the tool’s cost tier to the value of the output, and never let the expensive setting become your default. Generating a hundred exploratory drafts at max quality on the slow model is how a $9 idea turns into a real invoice. Draft cheap, finish expensive, and the split does exactly the budgeting for you if you let it. If you are choosing between several image engines rather than committing to one, the same logic applies to platforms like Krea’s multi-model studio: the cheapest per-image price is meaningless until you know how many retries you burn to get a keeper.

The limits worth knowing before you rely on it

Two honest caveats. First, resolution. The API accepts custom sizes up to a 3,840-pixel edge, so 3840×2160 landscape and 2160×3840 portrait are on the menu, per the launch details. But OpenAI marks any output above roughly 3.69 million total pixels as experimental, which is its own way of warning you that big dimensions do not guarantee crisp small text or clean product labels. If a headline or a legal line has to be pixel-perfect, generate the image and set the type in a real design tool.

Second, edit drift is reduced, not eliminated. Even Sunburst can alter details you did not ask it to touch after several passes, so the practical advice from hands-on reviewers holds: make one meaningful edit per turn, then compare at full zoom before you continue. The model is more consistent than its predecessor, which I covered when ChatGPT Images 2.0 launched, but consistency is a spectrum and you are still the quality gate.

The bigger picture, and the reason this release is worth your attention even if you already have an image tool you like: the whole category is drifting toward giving creators explicit controls instead of a single magic button. Two models with a real tradeoff, a sketch canvas for layout, editing that respects what you locked. Set against the consent-default problems that dogged tools like Meta’s Muse Image, a release whose headline feature is “here are more knobs, you decide” is a healthier direction. Use Flare to think, use Sunburst to finish, and draw the layout when words are failing you.

Ty Sutherland

Ty Sutherland is the Chief Editor of Full-stack Creators. Ty is lifelong creator who's journey began with recording music at the tender age of 12 and crafting video content during his high school years. This passion for storytelling led him to the University of Regina's film faculty, where he honed his craft. Post-university, Ty transitioned into the technology realm, amassing 25 years of experience in coding and systems administration. His tenure at Electronic Arts provided a deep dive into the entertainment and game development sectors. As the GM of a data center and later the COO of WTFast, Ty's focus sharpened on product strategy, intertwining it with marketing and community-building, particularly within the gaming community. Outside of his professional pursuits, Ty remains an enthusiastic content creator. He's deeply intrigued by AI's potential in augmenting individual skill sets, enabling them to unleash their innate talents. At Full-stack Creators, Ty's mission is clear: to impart the wealth of knowledge he's gathered over the years, assisting creators across all mediums and genres in their artistic endeavors.

Recent Posts