Skip to content
Grok image
English
English简体中文繁體中文DeutschEspañolPortuguês

GPT Image 2.5 Sunburst

GPT Image 2.5 Sunburst is the precision half of the pair OpenAI shipped on 8 September 2026. It sits next to GPT Image 2.5 Flare on this site, and the two pages are deliberately split: Flare's page covers the model most requests should use, and this one covers the tier you pick when an edit has to hold. One thing belongs before anything else, because it decides how the rest of the page reads — the two tiers have identical entries in the catalogue that drives this workspace, so nothing in the interface will tell you which one you are running.

Updated:

GPT Image 2.5 Sunburst

Choose a mode, write the prompt, then read the quoted cost before you generate.

AI Image0

Generated content will appear here

Third-party cases

Careful-work cases from third-party pages

Eleven images of the kind this tier is positioned for: product photography, packaging, brand layout, a campaign poster and two multi-frame sets. Together they carry the commercial half of the case list.

Every image here is third-party material retrieved on 19 September 2026, and none of it was generated by this site. The pages these came from describe GPT Image 2.5 as a whole and do not say which of its two tiers produced any of the images they show, so nothing in this gallery is attributed to Sunburst. Read it as an indication of the kind of work the release handles, not as evidence about the difference between the tiers. The prompt beside each image was written by this site from the picture, because the source pages publish no prompts, and the prompts are kept in English on every language version because changing the language would change the input the model receives.

A clear glass perfume bottle with amber liquid standing on a pale stone block against a warm neutral background Read the prompt ↗

01 / Third-party case, product photography

Amber glass on pale stone

weshop.ai ↗

Prompt for this image

A clear glass perfume bottle with amber liquid standing on a pale travertine block, seamless warm neutral background, soft studio light from the left, shallow depth of field.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

A white cylindrical air purifier standing on a wooden side table in a bright living room, a plant and a sofa behind it Read the prompt ↗

02 / Third-party case, product photography

Appliance on a side table

weshop.ai ↗

Prompt for this image

A white cylindrical air purifier on a light oak side table in a sunlit living room, a plant and a low sofa softly out of focus behind it, natural window light.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

An open book and a dark ceramic mug on a wooden table beside a window, with a folded cloth in the foreground Read the prompt ↗

03 / Third-party case, lifestyle

An open book and a mug

weshop.ai ↗

Prompt for this image

An open book and a dark ceramic mug on a warm wooden table beside a window, a folded cloth in the near foreground, soft daylight from the window and a blurred room behind.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

A black leather handbag with twin top handles and a slim front pocket, centred against a dark grey backdrop Read the prompt ↗

04 / Third-party case, product photography

A structured leather bag

weshop.ai ↗

Prompt for this image

A black leather handbag with two top handles and a slim front pocket, centred on a dark grey seamless backdrop, even studio lighting, the leather grain and stitching clearly legible.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

A woman in a dark blazer seated on a chair in a bright office, one hand resting on her knee, looking at the camera Read the prompt ↗

05 / Third-party case, business portrait

Seated in an office

weshop.ai ↗

Prompt for this image

A woman in a dark blazer seated on a chair in a bright modern office, one hand resting on her knee, looking directly at the camera, large windows behind her.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

A pale blue poster reading REFILL, REUSE, REPEAT in large letters, with a white reusable bottle standing at the bottom Read the prompt ↗

06 / Third-party case, text rendering

A bottle campaign with a headline

weshop.ai ↗

Prompt for this image

A pale blue campaign poster with the headline REFILL, REUSE, REPEAT in large capital letters, the spelling exact and the letters evenly spaced, a white reusable bottle standing centred at the bottom of the frame and no other text.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

An elevated night view of a wet city street, reflections across the roadway and lit windows on both sides Read the prompt ↗

07 / Third-party case, environment

Wet street from above

image25.io ↗

Prompt for this image

An elevated view along a rain-soaked city street at night, light reflecting across the whole roadway, lit windows and shopfronts on both sides, deep perspective.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

A dark presentation board holding several framed images and a small stack of cards, lit from one side Read the prompt ↗

08 / Third-party case, layout

A board of frames

image25.io ↗

Prompt for this image

A dark presentation board carrying several framed images and a small stack of cards, arranged as a layout rather than a single picture, lit from one side against a dark wall.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

A frosted serum bottle with a dark pipette cap standing among pale stone blocks with cool blue light Read the prompt ↗

09 / Third-party case, product photography

Serum bottle on stone blocks

image25.io ↗

Prompt for this image

A frosted serum bottle with a dark pipette cap standing among pale stone blocks, cool blue accent light from behind, product centred and label facing the camera.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

A four-panel set of the same rainy street with a figure in a blue coat walking through it in each panel Read the prompt ↗

10 / Third-party case, multi-frame set

One street, four frames

image25.io ↗

Prompt for this image

Four panels of the same rainy city street, the same figure in a blue coat walking through each one at a different distance, consistent lighting and colour across all four frames.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

Two panels showing the same white trainer on a plain grey surface, once from the side and once from above Read the prompt ↗

11 / Third-party case, comparison

The same shoe twice

image25.io ↗

Prompt for this image

Two panels of the same white trainer on a plain grey surface, one from the side and one from above, the laces, stitching and sole identical in both.

Open this prompt in the workspace ↗

Written by this site from the source image, not by the publisher.

What GPT Image 2.5 Sunburst is, in the publisher's words

OpenAI released ChatGPT Images 2.5 on 8 September 2026 and shipped two image models in its API alongside it. The announcement describes the second of them like this: Sunburst is built for premium visual workflows that benefit from tighter control across edits, for work like production-ready campaign creative or polished product imagery, and it offers that extra level of precision with longer generation times.

Read against the same announcement's description of Flare — a higher-quality image than the previous generation at half the latency, and the default choice for most applications — the split is not bigger versus smaller. It is faster versus finer, and the publisher pays for the finer one in time, not in money and not in a different set of controls.

The timeline claim is given for the family rather than for Sunburst alone. The announcement says image generation latency is down by up to 50 percent compared with the previous generation, and one design partner quoted in it reports the fast tier running at two to four times the speed of the older model. There is no published figure for how much slower Sunburst is; only the direction is stated.

The only official number that separates the two tiers

The publisher's system card for this release publishes a safety evaluation that breaks the two tiers apart by name, and it is the single place where a published figure compares them directly. The measurement is unsafe output presented — generations that escaped the safety stack — and lower is better.

Two rows deserve a second look, and both cut against a tidy story. Hate is identical on both tiers and on the previous generation to the hundredth of a percent, and wrongdoing is worse on both new tiers than on the model they replace. A page that quoted only the overall row would be telling you a simpler story than the table does.

This is a safety evaluation and not a picture-quality one. It is not evidence that Sunburst draws better, sharper or more accurate images, and this page does not use it that way. Its value is narrower and real: it is the only published figure that treats the two tiers as two different things rather than two settings, and it shows the slower tier escaping its own filters less often. The publisher also warns that sample sizes differ by policy, that the results come from a fixed adversarial set, and that they are not representative of everyday traffic.

CategorySunburstFlarePrevious generation
Overall1.09%1.41%1.64%
Sexual0.52%1.04%2.07%
Violence and gore0.64%1.93%1.93%
Political0.35%0.71%1.06%
Abuse1.69%2.54%2.54%
Nonviolent and violent wrongdoing2.86%3.27%2.45%
Hate0.36%0.36%0.36%
Extremism0.00%0.00%2.56%

The two catalogue entries are field for field identical

This site does not decide what a model can do. The workspace renders whatever the service catalogue declares for a product, and for these two products it declares the same thing twice. Read the two columns: there is no field where they differ.

This is the most useful thing this page can tell you, and it is also the least flattering. If you expected the precision tier to hand you a slider the fast tier does not have, it does not. Choosing Sunburst changes which model the request goes to and nothing about the interface. The precision, if it arrives, arrives in the picture.

It follows that the control named render quality is the dial worth understanding here, and that it is available on both tiers rather than being the thing that separates them. It defaults to low, which is worth knowing before judging a first result: a soft first image from either tier may be the dial rather than the model.

Declared fieldGPT Image 2.5 SunburstGPT Image 2.5 Flare
Aspect ratiosthe same sixteenthe same sixteen
Resolutions1k, 2k, 4k1k, 2k, 4k
Default resolution1k1k
Render qualitylow, medium, highlow, medium, high
Default render qualitylowlow
Reference imagesup to 10up to 10
Speed controlnot exposednot exposed

Three reference-image claims, and which one this site can verify

The three sources that describe this model do not agree on how many reference images it takes. Rather than pick the most convenient one, here they are side by side.

So the number this page can stand behind is ten, and it is a statement about this workspace rather than about the model. If a task needs more references than that, the honest thing is to test it here rather than trust either figure: the limit that matters is the one the service enforces when you submit.

SourceClaimCan this page verify it?
The publisherpublishes no reference-image limit in the announcement or the system cardno claim to check
This workspace's catalogueup to 10yes, it is the entry the workspace runs on
A third-party product page ranking for this nameup to 16no source is given for it

Editing without drift: what the publisher says changed

The precision claim has a concrete form, and the announcement states it in the same passage as the editing behaviour: the model produces more consistent results when editing images, and users can change an image's setting, style or composition while retaining more of the details. The system card adds that sketch-based generation, templates and shared prompts exist to give users more control over the process.

Multi-turn editing is the case that matters most and the case least supported by published material. Every edit is a fresh request, so small drifts accumulate across a sequence: a face shifts, a logo loses a letter, a colour slides. The publisher's claim is that this tier drifts less. Nothing in the announcement quantifies it, and the demonstration material for it ships as video with no downloadable file, so this page cannot show you the comparison.

What you can do instead is make drift visible on your own task. Give the model one reference image of the subject, change exactly one thing per turn, and compare the first and the last turn side by side. If the subject is still recognisably the same, the tier is doing what it was bought for.

Editing jobs, with prompts

These four prompts are written by this site for the jobs this tier is positioned for. They are kept in English on every language version of this page, because the prompt is the input the model actually receives and translating it would change the result.

Change the background, keep the subject
Keep the product exactly as it is, including its label, its proportions and the direction of the light. Replace the background with a plain warm-grey studio sweep and add a soft shadow falling to the right.
Hold a character across two turns
Use the reference image for the face and hair only. Put the same person in a rain-soaked street at night, same expression, same jacket, and do not change the shape of the face.
Editable text in a layout
A campaign poster for a bottle of mineral water, headline reading REFILL REUSE REPEAT in tall condensed capitals, the spelling exact, the bottle in the lower third and no other text anywhere in the frame.
Change one area only
Change only the surface the handbag is standing on, from pale stone to dark wood. Leave the bag, its hardware, the background and the lighting untouched.

What the publisher does not publish

Four things a reader would reasonably expect to find are absent from both official documents, and saying so is more useful than filling the gaps with a third party's numbers.

Everything on this page attributed to the publisher can be read at the announcement and the system card. Everything attributed to this workspace is in the catalogue the workspace runs on. Where neither applies, the page says so instead of borrowing authority.

  • Resolution and aspect ratio: neither the announcement nor the system card lists supported output sizes. The sixteen ratios and three resolution tiers in the table above come from this workspace's catalogue, not from the publisher.
  • A reference-image limit: also absent, which is why the table earlier on this page has three rows.
  • A generation time for Sunburst: the announcement says only that it is longer, and the family-level figure is a 50 percent reduction in latency compared with the older generation.
  • The demonstration material itself: the published examples of precise editing and multi-turn consistency are video players on the announcement with no downloadable files, so they cannot be embedded or checked frame by frame here.

How a task runs here

Six steps, and the sixth is the one that matters for this tier.

  1. Open the image workspace and choose GPT Image 2.5 Sunburst from the model list.
  2. Write the prompt, or start from an image if you are editing.
  3. Attach reference images if the result has to follow something. The workspace allows up to ten.
  4. Set the aspect ratio, the resolution and the render quality. The defaults are auto, 1k and low.
  5. Submit. The server quotes the task for those settings before anything is charged.
  6. Open the result again and make the next edit from it rather than from a fresh prompt. This tier is bought for the second and third turn, not the first.

Where these cases came from, and what they cannot prove

Every image below comes from one of three third-party pages that rank for this model's name. None of them is official, and here is the part that had to be checked carefully: none of those three pages states which of the two tiers produced the images it shows.

This site's usual practice is to prefer the publisher's own material. It was not available for this page. The announcement's photographic figures were all used on the sibling Flare page, its editing and multi-turn demonstrations publish no image files at all, and the system card contains no pictures. What remained was the third-party set you see here.

Two consequences follow, and both are stated rather than smoothed over. The cases show product photography, packaging, brand layout and text rendering across a range of styles, which is the kind of work this tier is meant for, so they are relevant. And they cannot tell you whether Sunburst or Flare made them, so they prove nothing about the difference between the two tiers. Treat the gallery as an indication of what the release can do in careful hands, not as a comparison.

The prompts printed beside each image are written by this site from the picture, not by the model that made it. The source pages publish no prompts.

Choosing between the tiers when the controls do not decide it

Since the interface is identical, the choice has to be made on the work itself. Four questions settle most cases.

One caution about the last point. Because the two catalogue entries are identical, switching tiers mid-task changes nothing you can see in the settings and everything about the result. If you are comparing them, change one tier and nothing else.

  • Will there be a second edit? A single image with no follow-up is a job for the fast tier; a sequence where turn three has to match turn one is what this one is for.
  • Is there text, a logo or a label that has to survive? Product packaging and poster layouts are where drift is most visible and least forgivable.
  • Is the image going into a campaign or a catalogue rather than a feed? The publisher positions this tier at exactly that, and the longer generation time is the price.
  • Are you still deciding what the image should be? Then neither the tier nor the render-quality dial matters yet. Use the fast tier at low quality to find the idea, and spend the precision tier on executing it.

What this page does not claim

Four statements were left out on purpose, and each is a place where a more confident page would have been a less accurate one.

This page does not claim Sunburst produces better-looking images than Flare. The publisher positions it for tighter control across edits, and the one published figure that separates them is a safety measurement, not a quality one.

It does not quantify the time difference. The publisher states the direction and gives a family-level latency figure only.

It does not repeat a reference-image limit of sixteen. That number appears on a third-party page with no source, and the workspace enforces a different one.

No price is stated. The server quote returned when a task is submitted is the number that is billed.

Examples

A woman standing on a balcony at night in neon light, a dense lit city behind her

A portrait set against a neon city, published on a third-party page ranking for this model's name. It is here for the lighting: several coloured sources on one face, which is where an edit stops matching the frame it was made in.

a2e.ai, product page for GPT Image 2.5, retrieved 2026-09-19. Third-party material; the page does not state which tier produced it. · Third-party material, not official. Rights remain with the publisher of that page; shown with attribution and no claim of ownership.

An illustrated portrait in a Chinese opera style, with flowers and soft watercolour colour

An illustrated portrait in a traditional opera idiom. Included as a style case: the question here is whether a specific decorative language survives an edit, not whether the picture is pretty.

a2e.ai, product page for GPT Image 2.5, retrieved 2026-09-19. Third-party material; the page does not state which tier produced it. · Third-party material, not official. Rights remain with the publisher of that page; shown with attribution and no claim of ownership.

A huge fish inside a clear plastic bag held above a city street, the words OUT OF SCALE across the top

A surreal composite with lettering built into the composition. Two jobs in one image: holding an impossible scale relationship together, and getting a short phrase into the frame without it drifting into nonsense.

a2e.ai, product page for GPT Image 2.5, retrieved 2026-09-19. Third-party material; the page does not state which tier produced it. · Third-party material, not official. Rights remain with the publisher of that page; shown with attribution and no claim of ownership.

A magazine cover mock-up with ISSUE NO.5 and FULL MOON in large type over a moon illustration

A magazine cover mock-up. Cover lines are the hardest text case in image generation because every line has to stay where it was put, in the right order, at a readable size.

a2e.ai, product page for GPT Image 2.5, retrieved 2026-09-19. Third-party material; the page does not state which tier produced it. · Third-party material, not official. Rights remain with the publisher of that page; shown with attribution and no claim of ownership.

A retro food poster with GUANGZHOU in tall letters above three ripe tomatoes

A retro food poster, the second text-heavy case. The useful test in this one is the lettering treatment itself: a stylised face with a chipped edge breaks under an edit far more easily than a clean sans-serif.

a2e.ai, product page for GPT Image 2.5, retrieved 2026-09-19. Third-party material; the page does not state which tier produced it. · Third-party material, not official. Rights remain with the publisher of that page; shown with attribution and no claim of ownership.

An embroidered panel depicting a city map with buildings and water in thread, on a pale ground

An embroidered map, kept as a material case. Thread, stitched edges and a woven ground are fine repeating texture, which is the detail that softens first when an image is resized or re-edited.

a2e.ai, product page for GPT Image 2.5, retrieved 2026-09-19. Third-party material; the page does not state which tier produced it. · Third-party material, not official. Rights remain with the publisher of that page; shown with attribution and no claim of ownership.

Four hand-written Chinese characters on a plain white page, in a loose brushed hand

Four brushed characters on white. This is the strictest text case in the set: a script that has to look written rather than set, with stroke weights that a wrong edit turns into decoration.

a2e.ai, product page for GPT Image 2.5, retrieved 2026-09-19. Third-party material; the page does not state which tier produced it. · Third-party material, not official. Rights remain with the publisher of that page; shown with attribution and no claim of ownership.

An impressionist cityscape of a waterfront skyline in pastel colour, with Chinese characters in the upper left

A pastel impressionist skyline with a short caption in the corner. Two disciplines in one frame: a loose painted surface, and a few characters that still have to be the right characters.

a2e.ai, product page for GPT Image 2.5, retrieved 2026-09-19. Third-party material; the page does not state which tier produced it. · Third-party material, not official. Rights remain with the publisher of that page; shown with attribution and no claim of ownership.

Questions

Is GPT Image 2.5 Sunburst the better of the two tiers?

Not in the sense of drawing better pictures. The publisher describes it as built for premium visual workflows that benefit from tighter control across edits, and prices that in generation time rather than quality. The one published figure comparing the two tiers directly is a safety measurement, and this page does not stretch it into a quality claim.

If I pick Sunburst, will the workspace show me different settings?

No. The catalogue that drives this workspace declares the same fields for both tiers, field for field: the same sixteen aspect ratios, the same three resolution tiers, the same three render-quality settings and the same ceiling of ten reference images. Choosing this tier changes which model the request goes to and nothing in the interface.

How many reference images can I attach?

This workspace's catalogue declares a ceiling of ten. A third-party product page ranking for this name says sixteen without giving a source, and the publisher publishes no limit in either official document. The number that matters is the one the service enforces when you submit the task.

What does the render-quality setting do?

It offers low, medium and high, and it defaults to low on both GPT Image 2.5 tiers. Because it defaults low, a soft first result may be the dial rather than the model. It is worth knowing that this control exists on both tiers, so it is not what separates Sunburst from Flare.

How much slower is Sunburst than Flare?

The publisher does not say. The announcement states that Sunburst offers its extra precision with longer generation times, and gives a family-level figure of up to 50 percent lower latency than the previous generation. One design partner quoted in it puts the fast tier at two to four times the speed of the older model. There is no published number for the gap between the two tiers.

Can I see the published examples of precise editing?

Not on this page. The publisher's demonstrations of precise editing and multi-turn consistency are video elements on the announcement with no downloadable files, so they cannot be embedded or inspected frame by frame here. The cases on this page are third-party stills instead, and every one of them says so.

Do the cases on this page prove Sunburst is more precise than Flare?

No. Every image here comes from a third-party page that describes GPT Image 2.5 as a whole, and none of those pages says which tier produced the images it shows. They indicate the kind of work the release handles; they say nothing about the difference between the tiers.

Is this the same model as the GPT Image 2.5 Flare page on this site?

They are two products from one release on 8 September 2026, and this site keeps them on separate pages on purpose. The Flare page covers the tier the publisher calls the default choice for most applications and holds the main keyword for the model. This page covers the precision tier and the editing work it is meant for. Both link to each other.

Explore next