GPT Image 2.5: Flare or Sunburst, what actually changes, and when it lands on Imaginode

this time, the model has actually shipped

ChatGPT Images 2.5 has been live since 8 September 2026, and the OpenAI API lists two models dated the same day: gpt-image-2.5-flare and gpt-image-2.5-sunburst.

At the end of August, we were explaining that Nano Banana 3 did not exist, and that the two anonymous models tested on LMArena in August, mona-lisa-1 and luna-lisa-alpha, smelled more of OpenAI than of Google. Well, there it is, the answer came this Tuesday 8 September. OpenAI has published ChatGPT Images 2.5, available in ChatGPT for every account, free ones included, in ChatGPT Work and in Codex, on desktop, mobile and web. Nothing officially proves that the two arena aliases were this model, but the timing matches to the day, and nobody at OpenAI has bothered to deny it.

On the developer side, the spec sheet is clear, and that is the one that matters for us. The API changelog says, under 8 September 2026: release of GPT Image 2.5 Sunburst and GPT Image 2.5 Flare for image generation and editing, through the Images API and the image generation tool of the Responses API. Both identifiers point to dated versions, gpt-image-2.5-flare-2026-09-08 and gpt-image-2.5-sunburst-2026-09-08. Two names, two models, one shared price list, we come back to that further down.

This article does what we do at every release: it separates what OpenAI announces from what the first users measure, it turns the token price list into cents per image, it settles Flare versus Sunburst, and it tells you where to use it tonight. And it answers the question we have already been asked three times today: no, GPT Image 2.5 is not on Imaginode yet, and here is why, with no corporate spin.

what changes, according to OpenAI and according to the first feedback

OpenAI announces latency cut by up to 50% compared with Images 2.0, more natural light, richer textures, better fidelity to reference photos and successive edits that no longer drift.

The first number is speed. OpenAI says it has cut generation latency by up to 50% compared with Images 2.0, and presents Flare as delivering better quality images than GPT Image 2 with half the wait. The Manus evaluation team, quoted in the announcement, goes further: in its tests, Flare produces its images two to four times faster than GPT Image 2, with transparent backgrounds that are finally clean thrown in, which matters to anyone making logos and interface elements.

The second axis is editing. The model is supposed to change only what you ask for, preserving the subject, the composition and the background, and above all to hold its edits across several turns of conversation. That is exactly the point we flagged as the next front in the Nano Banana article: the eighth edit that warps the face you approved at the second. If OpenAI has genuinely fixed that, it is more useful than a gain in raw beauty on a single isolated image. Transparent backgrounds and busy layouts, posters and visuals with several blocks of text, are announced as better handled too.

Now the honest caveat. OpenAI has published no benchmark table with numbers, no arena ranking, and the only independent feedback of the day comes from a partner quoted by OpenAI. We have not yet been able to run our own test boards, the same ones we use for every model in the catalogue, since the model is not yet where we plug our models in. The sentences above are therefore a vendor's promises, credible given the lineage, not measurements. Bear in mind that the difference between the two usually shows on the third day, not the first.

Flare or Sunburst: which one to ask for

Flare is the default choice, fast and built for volume; Sunburst is presented as the more capable of the two, cut for editing precision, with longer generation times.

OpenAI describes Flare as the everyday model: fast, good quality generation, designed for creator and social media content, product experiences, visual search, prototyping and high volume generation. Sunburst's sheet, for its part, says "our most capable model for image generation and editing", with the emphasis on edit precision, and warns of longer generation times. The official guide recommends it for the pipelines where editing precision matters most: campaign visuals ready to deliver, polished product photos.

Granted, on paper, everyone wants the more capable one. But look at the price list before choosing: both models are billed at exactly the same token rates. The difference between Flare and Sunburst is therefore not a difference in money, at least for now, it is a difference in waiting time and, if the promises hold, in fidelity on fine edits. Which gives a simple rule.

Personally, I would take Flare for everything that is still being explored: the thirty variants of a YouTube thumbnail, framing attempts, the first versions of a visual. And I would keep Sunburst for the moment when the image is approved and there are three precise edits left to make on a product photo that is going to a client. Generate fast to decide, generate slowly to deliver: it is the same logic as Nano Banana versus Nano Banana Pro, and it has never let us down.

the price: the same tokens as GPT Image 2, but two extra quality tiers

Both models cost $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens, the exact GPT Image 2 price list; the new part is the xhigh and max qualities above high.

The official price list is identical for Flare and Sunburst, and identical to that of GPT Image 2: $5 per million text input tokens ($1.25 cached), $8 per million image input tokens ($2 cached), $30 per million image output tokens. Flare's sheet says it in black and white, the token rates match those of GPT Image 2. If you read elsewhere that the price has doubled, that person is comparing with an old price list; OpenAI's pricing page, as of the evening of 8 September, shows the same numbers on both lines.

The real pricing news is elsewhere, and it is easy to miss. GPT Image 2 offered three qualities, low, medium and high. Both 2.5 models accept six: low, medium, high, xhigh, max and auto. An image is paid for by the number of tokens it contains, and a higher quality tier means more tokens for the same size. In other words, xhigh and max will cost more than high, and OpenAI itself warns that its GPT Image 2 cost calculator does not estimate the token consumption of 2.5. Nobody, tonight, can give you the exact price of an image in max. We will measure it with a real request, as we do for every model.

For context, here is what GPT Image 2 costs today on Imaginode, with a credit worth one cent: 1 credit per image in low quality, 8 credits in medium, 30 credits in high. Those three tiers will not move by a cent with the arrival of 2.5, since the tokens cost the same price. What will need watching is the two new floors. A max tier on a social media image seen for three seconds on a phone is using a bazooka to swat a fly, and a fine way to throw your credits out of the window. Keep it for what will be printed or blown up.

A word on sizes, because the question comes up every time. The API accepts three recommended formats, 1024 x 1024, 1536 x 1024 and 1024 x 1536, and free dimensions in width x height as long as each side is a multiple of 16, the ratio stays between 1:3 and 3:1 and no edge exceeds 3840 pixels. PNG output by default, JPEG or WebP with a compression setting, transparent background on request. If you are preparing visuals for several platforms, the guide on formats and resolutions per network gives the exact sizes to ask for.

Balance1 250

Flux Schnell1 cr

Flux 2 Pro7 cr

Kling 2.5 · 5 s · 720p26 cr

Seedance 2.0 · 5 s · 4K330 cr

Generate · 26 cr

Sketch, comments, templates: what ChatGPT adds, and what it is worth

In ChatGPT, Images 2.5 arrives with a sketching tool called with @Sketch, comments placed directly on an area of the image, format templates and an editing bar with version history.

On the interface side, OpenAI delivered four things alongside the model. Sketch, first: you type @Sketch in the conversation, you draw a freehand sketch, you add a style description, and the model uses it as a composition guide for the final image. Comments, next: instead of rewriting the whole prompt, you point at an area and write "make the sky more dramatic", and only that area moves. Templates, finally, for poster, flyer, merchandise and product photo, which ask structured questions before generating. And an editing toolbar: annotation, comment, background removal, eraser, resizing, with version history kept.

I will be blunt: it is rather very good, and it is not new. A sketch that guides the composition is a reference image plugged into the model's input. A comment on an area is a localised edit with a mask. A template is a structured prompt prepared in advance. They are, one by one, the building blocks of a workflow canvas, what we call nodes, and what equipped creators have been doing for a year with Flux Kontext, Nano Banana or GPT Image 2. OpenAI has just put those gestures into a chat window, which is excellent news for the hundred million people who will never open a canvas.

What the chat window still lacks, and it is structural, is repeatability. A comment on an image is done once. A chain of nodes, reference, generation, localised edit, upscale, replays across the fifty products of a catalogue by changing a single input. If you produce three visuals a month, ChatGPT with Sketch will be more than enough. If you produce three a day, the tool that counts is not the one that draws most prettily, it is the one that does the same thing again without asking you to start over. The article on reproducing the same image explains the difference in practice.

where to use it tonight, and why not yet on Imaginode

Three doors open on day one: ChatGPT for every account, the OpenAI API for anyone managing a key, and Adobe Firefly; the Vercel AI Gateway catalogue, through which Imaginode wires up its models, does not list GPT Image 2.5 yet as of the evening of 8 September.

The shortest door is ChatGPT. The model is rolled out to every account, free ones included, on desktop, mobile and web, and the generation quotas remain those of each plan, which OpenAI still does not publish. To form an opinion in ten minutes and try Sketch, that is where to go. Our comparison with ChatGPT details what those unwritten limits mean when you are producing for real. Second door, the API, with the two identifiers above, a key, usage billing, and the option of choosing the quality and the size to the pixel. Third door, Adobe Firefly, which is integrating it into its studio from today.

And Imaginode, then? We wire up our image and video models through the Vercel AI Gateway catalogue, which lets us have Flux, Seedream, the three Nano Bananas, Kling or Veo behind a single credit counter, with a price displayed before every generation. That catalogue, checked tonight, contains four OpenAI image models: GPT Image 1, GPT Image 1 Mini, GPT Image 1.5 and GPT Image 2. No 2.5 line, neither Flare nor Sunburst. As long as the line is not there, the model is not with us, and I would rather tell you than let you hunt for a button that does not exist.

What that changes for you: nothing to do on your side. The day the Gateway lists gpt-image-2.5-flare and gpt-image-2.5-sunburst, we measure the real costs per tier with a real request, we set the price in credits, and the model appears in the list, in the same canvas, next to GPT Image 2 which stays where it is. No OpenAI subscription to take out, no key to manage. This article will be updated that day with the observed prices, and the models page is the authoritative source in the meantime.

Imaginode Assistant from 1 credit / message

I want a 10-second ad for my coffee brand ☕

Here's the workflow I suggest 👇

Add this workflow to the canvas

in the meantime, which model for which job

GPT Image 2 at 8 credits remains the most reliable in the catalogue for text inside the image, Nano Banana 2 at 10 credits for editing that respects faces, Seedream 5 Lite at 5 credits for budget photorealism, and GPT Image Mini at 2 credits for exploring.

The September reflex would be to wait. That is a mistake, because the four models that cover GPT Image 2.5's ground today are already in the catalogue, and none of them is going away. GPT Image 2, at 8 credits in medium quality, is the one you reach for when there is text to write inside the image: an exact title, a short sentence, accented characters. It is the most obedient in the catalogue on that point, and it edits your input images. What 2.5 promises on top is to hold those edits over time and to go faster; what 2 already does is the job.

Nano Banana 2 at 10 credits is the direct competitor, and on retouching an existing photo, removing an object, changing an outfit, keeping the face intact, it stays ahead. For photorealism on a small budget, Seedream 5 Lite at 5 credits an image does things GPT Image does not do at that price, with characters that stay consistent from one image to the next. And for exploring, trying, getting it wrong ten times, GPT Image Mini at 2 credits does exactly what Flare will do on drafts: speed, for cheap.

The method does not change, whatever model ships next month: frame the idea on the budget model, validate the composition, and only move to the top of the range for the final version. Across ten visuals, the difference is counted in dollars. The calculator gives you the cost of a series before you launch it, and the per model pricing detail is public, tier by tier.

the model version is still not your limiting factor

What GPT Image 2.5 promises to do better, edits that no longer drift and a reference that is respected, is exactly what a node workflow guarantees by construction, whatever model is plugged into it.

Reread the list of announced improvements: the subject of the reference photo better preserved, the edit that touches only the requested area, the previous changes that hold at the next turn. Those are three flaws of the chat window, where every message puts everything back in play. In a canvas, the reference is a node that does not move, the edit is a node that receives the approved image and nothing else, and the previous version stays on the table, next to the new one. A model that drifts less is of course good news. A method where it cannot drift is better, and it already works with this summer's model.

The second lever, we repeat it because it pays more than a version change: the prompt. A three-word prompt gives a forgettable image on 2.5 just as it will on any model of 2027. A prompt that describes the framing, the light and the style, with a reference image to lock the subject, produces a deliverable image on GPT Image 2 today. The guide on writing a prompt that works takes ten minutes to read, and the one on deformed hands and illegible text fixes the two most frequent complaints, without waiting for anyone.

So yes, test GPT Image 2.5 in ChatGPT tonight, look at Sketch, form your own opinion on Flare. And if you see the line appear in the Gateway catalogue before we do, write to us, we like losing that kind of race. But the best use of these few days or weeks of waiting is putting in place the habits that will make this model useful the day it arrives: a locked reference, a complete prompt, a chain of nodes you replay. Those who have them will gain a full generation. Everyone else will make slightly prettier images, slightly faster, out of three-word prompts.