GPT Image 2.5 Flare is one of two image models OpenAI released in the API alongside ChatGPT Images 2.5. This page keeps three kinds of statement apart, because they have different evidence behind them. The first is what the publisher said, quoted and attributed. The second is what this site's own catalogue returns for the product, which is what the workspace actually obeys. The third is what the published artwork can and cannot support. The page carries eleven cases, and says plainly where that falls short of the usual fifteen.
Updated:
GPT Image 2.5 Flare
Choose a mode, write the prompt, then read the quoted cost before you generate.
Seven pieces from OpenAI's own announcement. Every one of them is the publisher's material, retrieved on 19 September 2026 and stored on this site's own media host so the page does not depend on someone else's uptime.
These images were published with the ChatGPT Images 2.5 announcement. The announcement does not say which of the two API models produced any of them, and this page will not invent that attribution. Each prompt beside an image was written by this site from the finished picture, not published by the publisher, and is kept in English because changing its language would change what the model receives.
A rain-slicked elevated walkway in a dense future city at night, a single small figure silhouetted at the left, enormous pale faces rendered on the building facades, a lit suspension bridge spanning the street behind them, wet reflections on every surface, cold blue-grey light with warm window accents, wide cinematic composition.
A studio portrait in the style of an early nineteen eighties album sleeve, a smiling man with an afro wearing a bright colour-blocked windbreaker and a gold chain, magenta and cyan light bars raking across a dark background, a portable cassette radio on the right, warm key light on the face, square crop at chest height.
A mosaic artwork in blue and gold glass tiles, showing the curved horizon of Earth across the lower half with continents outlined in light tiles, an expanse of stars above with planets and a spiral galaxy, every element built from the same square tile grid with visible grout lines, deep navy background, decorative and symmetrical.
A retro-futurist illustration of a family of three seen from behind at a large curved window, looking out over a green valley with a lake, winding roads and low white buildings, the landscape curving upward into a vast cylindrical habitat, a small spacecraft and a moon in the dark sky to the left, warm interior light against cool exterior, nineteen fifties magazine painting style.
A surreal science fiction scene with a dense city of lit towers hanging upside down from the top edge of the frame, its buildings pointing downward, below it an open dark sky filled with stars and a thin layer of cloud, one small figure floating weightless in the middle distance, deep blue palette with warm orange window lights, very wide composition.
A sheet of eight illustrated stamps in two rows of four, each showing a different American national park landscape in a vintage poster style with a white perforated border, legible labels under each scene reading Yellowstone, Grand Canyon, Acadia, Zion, Glacier, Great Smoky Mountains, Denali and Everglades with their states, muted reds greens and blues, flat even lighting.
A photograph of a printed studio portrait held in one hand, showing a seated child in a red sweatshirt against a mottled blue studio backdrop, the paper slightly curled, a cluttered domestic room visible around and behind the print, flat even scan-like lighting, portrait orientation.
Written by this site from the published image, not by the publisher.
What GPT Image 2.5 Flare is, in the publisher's words
OpenAI announced ChatGPT Images 2.5 on 8 September 2026 and, in the same announcement, released two image models in its API. Of Flare the announcement says it brings the same improvements in quality, editing, and speed to the API and is the default choice for most applications, delivering higher-quality images than GPT-Image-2 at 50% lower latency. The same announcement describes Sunburst as built for premium visual workflows that benefit from tighter control across edits, with longer generation times.
The announcement lists where Flare is meant to be used: creator and social content, product experiences, visual search, rapid image prototyping, and high-volume generation. Two claims in that list are numeric and worth separating from the rest. The quality comparison is against GPT-Image-2 specifically, not against Sunburst or against anything outside this family. The latency figure is the publisher's own measurement, published without a test method, so this page repeats it as the publisher's claim and does not restate it as a measured result.
On safety, the announcement says the model builds on existing safeguards with checks on prompts and images, and that the publisher continues to use C2PA metadata and invisible watermarking to identify images made with its tools. That matters for anyone publishing output commercially, because it means provenance metadata may travel with the file regardless of what this site or any other reseller does.
Attribution note. The announcement page could not be read through a headless browser from this network, which returns a bot challenge instead of the article. The statements above were taken from the publisher's own announcement text and checked against a second official language version of the same page. Where this page quotes the publisher it says so; where it describes this catalogue it says that instead.
Flare against Sunburst: the split the publisher made
The two API models are not a smaller and a larger version of one thing. They are two answers to the same request, and the publisher's own framing is that one optimises for throughput while the other optimises for control. Both accept the same kinds of input and produce the same kind of output.
Every row in that table is the publisher's positioning, not a measured comparison, and this site has run no side-by-side test of its own. The two models cost the same here, so the choice is not a price decision.
Choose Flare when the number of acceptable attempts matters more than the quality of any single one. Its stated advantage is lower latency, which only pays off if you generate several candidates and keep one.
Choose Sunburst when a specific edit has to land and the extra time is affordable. Its stated advantage is control across edits, which is a different thing from being better at generating from scratch.
A workable division of labour follows from that: explore with Flare, then move the winning direction to Sunburst for the pass that has to be final.
What the publisher says
Flare
Sunburst
Position in the API
Default choice for most applications
Built for premium visual workflows
Stated trade
50% lower latency than GPT-Image-2, higher quality
Tighter control across edits, longer generation times
Named uses
Creator and social content, product experiences, visual search, rapid prototyping, high-volume generation
This site does not decide what a model can do. The workspace reads the product entry the service catalogue returns and offers only what that entry declares. The table below is that entry, as read on 19 September 2026.
Task model routing is the least obvious row. The workspace does not send every request to one identifier: text-to-image and the two image-in modes are dispatched to different task models on the service side. That is why a failure in one mode says nothing about the other two.
Declared field
Value here
Catalogue name
GPT Image 2.5 Flare, in the ChatGPT Image family
Availability
Available and callable
Accepted modes
Text to image, image to image, reference to image
Task model per mode
Text to image uses the bare id; the other two route through -i2i and -r2i variants
Input modalities
Text and image
Reference image input
Accepts a URL or base64 data
The canvas: what sizes actually come out
Three resolution tiers are offered, and each tier carries its own set of pixel presets. The presets are not scaled versions of one another, so the tier you pick decides which aspect ratios you can have, not only how large the file is.
One inconsistency worth knowing before you file a bug. The catalogue lists 1:3 and 3:1 among the model's supported aspect ratios, and both have presets at 2K and 4K, but the workspace aspect ratio control does not offer them as options. A ratio the model supports and the service can serve is therefore not reachable from the picker today.
The hard ceiling is an edge of 3840 pixels and a maximum aspect ratio of 3:1. Anything wider than three to one is not a supported shape at any tier.
Total pixel count has to land between 655,360 and 8,294,400, and edges are constrained to multiples of 16. A request that misses either bound is not a slightly wrong size, it is an unsupported one.
The 1K tier has no preset for 1:3 or 3:1, so those two shapes start at 2K. If a layout depends on a three-to-one crop, the smallest option available is a 2048-wide edge.
The 4K tier does not exceed the 3840 edge limit, which is why 4K square lands at 2880x2880 rather than a nominal 4096.
Tier
Aspect ratios with a published preset
Example preset
1K
13, the narrowest set
1:1 is 1024x1024, 16:9 is 1280x720
2K
15, including 1:3 and 3:1
1:1 is 2048x2048, 16:9 is 2048x1152
4K
15, capped by the maximum edge
1:1 is 2880x2880, 16:9 is 3840x2160
The four-point tier, against its siblings
Three image products from the same publisher sit on this shelf, and they are priced a point apart. That ladder is the clearest way to see what the catalogue thinks each generation is worth.
Read that column carefully. The catalogue marks these as fallback reference prices, which is the figure used when no more specific billing rule matches. The binding price is the quote the server returns for the mode and settings you actually chose, immediately before the task starts. Flare and Sunburst cost the same as each other, so the publisher's split between speed and control is not a price decision in this catalogue.
Product here
Fallback reference price per image
Position
GPT Image 2
2 points
Previous generation, still available
GPT Image 2.5
3 points
Same generation, plain entry
GPT Image 2.5 Flare
4 points
Same generation, the fast default
GPT Image 2.5 Sunburst
4 points
Same generation, the precision option
What the settings do not expose
A landing page that lists only what a model can do is half a page. These are the controls a user might reasonably look for and not find, with the reason each one is absent.
None of these is a defect this page is reporting on the publisher's behalf. They are what the catalogue declares, and the page exists so that nobody has to discover an absent control by hunting for it.
There is no quality switch. The catalogue declares a single value for generation quality on this product, auto, with no low, medium or high alternatives. The three-tier control that other pages here describe does not apply to Flare.
The catalogue publishes no reference-image ceiling for this product. The field is absent, so the workspace keeps the conservative limit of the sibling model rather than inventing a larger one. Treat the current limit as this site's safety setting, not as a specification the publisher or the catalogue has stated.
Image detail is not adjustable. There is no seed, no step count, no sampler, and no guidance scale in the declared form. The form accepts a prompt, optional reference images, an aspect ratio, a resolution tier, and nothing else.
Aspect ratios of 1:3 and 3:1 are supported and presetted but not selectable, as noted above.
There is no negative prompt field. Anything you want excluded has to be written into the prompt as an instruction.
How a task runs here
The workspace pins the model when you arrive from this page, so nothing else has to be chosen by hand. From there the sequence is fixed, and every step happens before you spend anything you cannot see.
Open this page in the workspace, which fixes the model and reads the current catalogue entry for it.
Write the instruction. Describe subject, framing, lighting and materials in plain sentences. Flare has no field for exclusions, so state them in the same sentence as the thing they apply to.
Attach reference images if the task needs them. The mode changes to image to image or reference to image depending on what you attach, and the task model changes with it.
Pick a resolution tier and an aspect ratio. Check the preset table if the shape is unusual, because the narrow tier has no preset for the extreme ratios.
Confirm the quote the server returns for those settings. That quote, not this page, is the price of the task.
Prompts for the jobs Flare is positioned for
The publisher names social content, product experiences and rapid prototyping first. These three prompts are written by this site for those jobs, in English, and kept in English on every language version of this page because the prompt is the input the model receives. Translating it would change the task, not the translation of the task.
These are this site's own prompts. The publisher has not published the prompts behind the artwork in the gallery above, and none of these is a reconstruction of one.
Product shot with a controlled highlight
A premium studio product photograph of one clear glass bottle of amber liquid on a pale limestone plinth, warm neutral backdrop, soft late-afternoon light entering from the left producing a single controlled highlight through the glass, a natural contact shadow on the stone, vertical composition with clean empty space above the bottle for later copy. No flowers, no ribbons, no extra bottles, no brand text, no labels, no watermark.
Social variation set
A top-down flat lay of a ceramic coffee cup and a folded linen napkin on a dark walnut table, morning window light from the upper right, three soft shadows, muted palette of cream, walnut and slate, square composition framed tightly so the cup sits at the lower left. Keep the surface texture visible and the highlight on the cup rim intact.
Layout with rendered text
A three by three grid of flat poster designs on a single sheet, each cell a different bold geometric composition in primary colours, with short legible captions reading Travel Farther, Create, Grow Together, Listen More, Find Your Balance, Drink More Water, Learn Something New, See Beauty, Choose Kindness, one caption per cell, all text spelled exactly as given, even spacing between cells, no overlapping letters.
What the artwork proves, and what this page does not claim
The gallery and the case list come from OpenAI's own announcement, retrieved on 19 September 2026. That provenance is unusual for this site, where most model pages depend on third-party material, and it comes with a limit that has to be stated rather than glossed over.
The announcement publishes these images as ChatGPT Images 2.5 artwork. It does not label which of the two API models produced any individual image, and no page the publisher controls states it per image. Attributing one of them to Flare specifically would be an invention, so every caption says only that the image was published with the announcement.
The images demonstrate that the family can render legible text, hold a consistent subject, and follow a tight composition. They do not demonstrate a latency claim, a failure mode, or a comparison against Sunburst, and this page uses them for nothing beyond the three things they show.
The announcement also references sample videos, but publishes no video file addresses. This page therefore carries no video, rather than a poster frame standing in for one.
One downloaded candidate was rejected during review because it is a screenshot of a third-party application rather than a model output. It is excluded from the manifest, and the rejection is recorded in the project evidence rather than quietly dropped.
This page is not an official product description and is not endorsed by OpenAI. It records what this site's catalogue returns and what the workspace does with it.
Examples
Before
After
The publisher's own before-and-after pair. The instruction was to change what the child is wearing without disturbing anything else, and the pose, the backdrop, the framing and the lighting all survive the edit. This is the clearest evidence on the page for the claim that the model changes what you asked for and leaves the rest alone, and it is one example rather than a measurement.
OpenAI, ChatGPT Images 2.5 announcement, before and after images, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Text rendering at nine short captions in one frame. The letters are crisp, spelled as intended, and correctly spaced in a grid, which is the failure mode most image models still show. Worth noting for anyone planning to use generated images as layout: the text here sits on flat colour fields, the easiest possible case.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A named style applied convincingly, down to the red bridge and the pastel terraces. Style adherence is the easiest capability to show and the least useful for judging a production workflow, so this case is here to document range rather than reliability.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A dense decorative layout with lettering, a monogram, an ornamental border and a small landscape illustration, photographed as a physical object rather than delivered as a flat sheet. The names are placeholder names. Physical staging is a good test because the model has to keep the print plausible under raking light.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A poster in a worn-print style, carrying the publisher's own product wordmark and Japanese headline copy. It is the only case here that ships a brand name inside the generated frame, and it is the publisher's own brand, which is why it is kept rather than rejected. Useful as an example of legible display type at poster scale.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
The widest composition in the set, and the one that shows the highest detail density. Architecture, reflections, distant lights and readable signage all hold together at 16:9, which is the shape most social and web placements actually want.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A square portrait with a specific period look, coloured light, and a foreground object the model had to place without cutting it awkwardly. Faces are the part of an image viewers judge fastest, which makes this the most demanding case in the set even though it looks the simplest.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
A single material rule applied across an entire frame. Every element is built from the same square tile, including the stars and the spiral galaxy. Holding one constraint across hundreds of small repeated units is where generated images usually break down, and here it holds.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Complex geometry rendered as an ordinary view. The habitat curves upward into the distance, and the landscape, roads and buildings all follow the same curvature, which is the part a viewer notices immediately when it is wrong.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
An inverted composition with a heavy reflection risk. Towers hang downward over a starfield with no support, and the scene stays readable rather than turning into noise, which is what makes it worth showing.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Eight distinct miniatures on one sheet, each with its own scene and its own legible label. Laying out many small independent panels with correct text in each is harder than one large scene, and this is the strongest evidence in the set that small text survives at print-like scale.
OpenAI, ChatGPT Images 2.5 announcement, retrieved 2026-09-19. · Third-party material. Rights remain with the publisher; shown with source attribution and no claim of ownership.
Questions
Is GPT Image 2.5 Flare the same model as GPT Image 2.5?
No. They are separate products in the catalogue with separate task models, separate entries and different fallback reference prices. This site lists both, and Flare is the one this page describes.
Flare or Sunburst?
Take Flare if you generate several candidates and keep one, because its stated advantage is latency. Take Sunburst if a specific edit has to land and you can afford the extra time, because its stated advantage is control across edits. Both cost four points here, so price does not separate them.
How many reference images can I attach?
The catalogue publishes no ceiling for this product, so the workspace keeps a conservative limit of ten rather than inventing a larger one. Treat that as this site's safety setting, not as a published specification.
Is there a low, medium or high quality setting?
No. The catalogue declares a single value, auto, for generation quality on this product. What you control is the resolution tier and the aspect ratio.
Why is there no 3:1 option in the aspect ratio picker?
The model supports 3:1 and 1:3, and both have presets at 2K and 4K, but the workspace picker does not list them today. Until that changes they are not reachable from the interface, so a three-to-one layout cannot be requested from this page.
What does a task cost?
The catalogue shows a fallback reference price of four points per image, which is what applies when no more specific billing rule matches. The price that binds you is the quote the server returns for your chosen mode and settings before the task starts.
Are the images on this page outputs of Flare?
Unknown, and the page does not pretend otherwise. They were published with the ChatGPT Images 2.5 announcement, which never says which of the two API models produced any individual image. They are the family's artwork, not a Flare-specific set.
Do generated images carry a watermark?
The publisher states that it uses C2PA metadata and invisible watermarking to identify images made with its tools. That is the publisher's standard, not a setting this site controls, so assume provenance metadata may be present and plan for it if you are publishing commercially.
Why are there fewer than fifteen cases?
Eleven is the honest maximum. Twelve official images were retrievable, one of them was rejected on review because it is a screenshot of a third-party application rather than a model output, and the remaining eleven are used once each. This page does not pad the count with unrelated material.