GPT-Image-2.5 Flare vs Sunburst: Which Variant to Use and When

For a while the test for an image model was whether it could make one striking picture. These days it's whether the model slots into a real product without drama, whether that's a creative tool, a store catalog, a campaign mockup, or any pipeline where regenerating a whole image over one changed detail is a waste. GPT-Image-2.5 from OpenAI is aimed at that gap. OpenAI released it on September 8, 2026 as its latest family for generating and editing images, and the API ships it in two variants named Flare and Sunburst.

Quick overview

The family follows ChatGPT Images 2.0 and the GPT-Image-2 generation of the API. Sharper pixels aren't the headline. The models are built to keep reference images intact, edit only the region you point at, stay consistent over several turns, follow detailed instructions, render lighting and textures more naturally and produce transparent-background assets.

OpenAI reports image latency cut by up to 50% versus Images 2.0. Flare is the sensible default for most applications, since it favors speed and still looks good. Sunburst runs slower but suits hard edits and final assets where small details must survive.

Early human-preference data is encouraging too. In the Arena snapshot from September 9, Sunburst took first place for text-to-image at 1421 ± 13 Elo, Flare was second at 1399 ± 13, and Image 2 came third at 1381 ± 4.

Both variants support:

• Quality levels from low to max, with xhigh and auto in between
• Custom sizes of up to 8,294,400 pixels in total
• No edge longer than 3840 pixels
• Transparency, delivered as PNG or WebP
• Both the /v1/images/generations and /v1/images/edits endpoints

Pricing doesn't separate them either. Text input costs $5 per million tokens, image input $8 and image output $30, the same for each.

Two models, one family

"GPT-Image-2.5" names the family, not one API model ID. In code you pick between gpt-image-2.5-flare and gpt-image-2.5-sunburst. The easiest way to think of them is as two operating modes of one generation: Flare is tuned for responsiveness and scale, Sunburst spends extra time on precision.

Factor	Flare	Sunburst
Main goal	Fast, high-quality generation	Maximum precision in generation and editing
Latency	Faster	Slower
Good fit	Social assets, prototypes, visual search, high-volume apps	Campaign creative, product hero images, demanding edits
Editing	Strong reference fidelity and multi-turn consistency	Strongest choice when exact details matter
Suggested role	Default model	Precision tier

A new product should usually start on Flare. Keep Sunburst for final campaign images, tricky localized edits and requests where a small change mustn't disturb anything else in the frame.

Quality controls also grew. Earlier GPT Image models topped out at high, while this family adds xhigh and max, which gives you finer control over the speed versus fidelity tradeoff.

Custom output sizes follow OpenAI's image-generation guide. Width and height both need to be multiples of 16, and anything above 2560×1440 is currently labeled experimental.

What actually changed in day-to-day work

Reference images survive better

Working from a reference has always meant a compromise. Ask the model to change too much and it may also alter the person, product, pose, proportions or background you wanted left alone. GPT-Image-2.5 keeps more of those details while changing only what you requested, which helps with:

• Swapping a product background
• Producing campaign variations around the same character
• Changing clothing while keeping identity and pose
• Building alternate compositions from a single source asset

OpenAI's own "Remixed baby portrait" demo changes the outfit but keeps the pose and the blue studio backdrop.

Edits stay local

A good editor knows the difference between "change the headline" and "rebuild the whole image." Picture a product shot with a bottle, a logo, a backdrop, shadows, props and promo text. The model is now better at touching one element without quietly bending the bottle's shape, the camera angle, the lighting or the surrounding layout. In commercial work that matters a lot, because a marketing team wants a controlled revision, not a lookalike replacement.

Longer editing sessions hold up

Real image work is iterative. A typical session: build a product scene, swap the backdrop, shift the product, rewrite the headline, add a promo badge, tweak the lighting, then export a transparent cutout. With weaker models each turn risks eroding what earlier turns established. This family is designed so previous edits persist across turns, which makes it feel like an interactive editor and less like a chain of unrelated generations.

Lighting and texture

OpenAI also reports more natural lighting and material textures. That shows up in skin, fabric, glass, metal, packaging, furniture, food and realistic environments, all places where a tiny artifact makes an image look obviously synthetic.

Briefs instead of captions

Image prompts now read like design briefs. One request might specify:

• Where several objects sit relative to each other
• The product package and the brand palette
• The camera angle
• What goes in front and what stays behind
• Light direction and mood
• Headline copy
• Areas that should be see-through
• Anything that must not appear

The new models turn that kind of structured instruction into a coherent image more reliably.

Transparency and sizes

Transparent backgrounds work with PNG or WebP output:

background="transparent"
output_format="png"

Any valid WIDTHxHEIGHT pair is accepted, not just a short list of preset aspect ratios. That suits isolated products, stickers, overlays, presentation graphics, UI components, thumbnails and ad assets.

Reading the benchmarks

Be careful with image benchmarks. A single score can't cover aesthetics, prompt adherence, how local an edit stays, realism, text rendering and reference fidelity all at once. For this release three kinds of evidence are worth a look: Arena human preferences, OpenAI's latency claims and the safety evaluation.

Arena: text-to-image

Arena runs blind comparisons. People see outputs from two anonymous models answering the same task and pick the image they like better. The preliminary September 9 snapshot looked like this:

Model	Text-to-image Elo	Rank	Votes
Sunburst	1421 ± 13	1	3,149
Flare	1399 ± 13	2	2,856
Image 2 (medium)	1381 ± 4	3	78,731
MAI Image 2.6	1331 ± 7	4	11,213
Grok Imagine Image 2.0 (low)	1315 ± 12	5	2,681

Both new variants sit ahead of Image 2 there. Still, this is an early preference signal, not a promise for every prompt or use case.

Arena: single-image editing

The distance grows on Arena's single-image editing board:

Model	Single-image edit Elo	Difference vs Image 2
Sunburst	1520 ± 9	+59
Flare	1491 ± 9	+30
Image 2	1461 ± 3	Baseline
Grok Imagine Image 2.0	1439 ± 8	-22
MAI Image 2.6	1434	-27

Sunburst's bigger lead in editing fits its job: it's the stronger option when precision and control outrank response time.

Latency

OpenAI says GPT-Image-2.5 lowers generation latency by up to 50% relative to Images 2.0. More specifically, Flare is pitched as delivering higher quality than Image 2 at half the latency. Sunburst makes the opposite trade on purpose and takes longer when the extra processing improves the final image or edit.

Safety

OpenAI published an adversarial safety evaluation, using prompts deliberately written to request policy-violating images:

Model	Safe generated	Unsafe blocked	Unsafe presented
Sunburst	77.0%	21.9%	1.09%
Flare	79.4%	19.2%	1.41%
ChatGPT Images 2.0 baseline	75.2%	23.1%	1.64%

Against the 1.64% baseline, the final unsafe rate is 0.55 percentage points lower for Sunburst and 0.23 points lower for Flare. These are stress-test numbers, not a measure of ordinary user traffic or of image quality.

Flare, Sunburst and Image 2 compared

Dimension	Flare	Sunburst	Image 2
Main goal	Speed plus high quality	Maximum precision	Previous-generation general image model
Text-to-image	Excellent	Best current Arena result	Excellent
Image editing	Excellent	Best of the three	Strong
Arena T2I Elo	1399 ± 13	1421 ± 13	1381 ± 4
Arena image-edit Elo	1491 ± 9	1520 ± 9	1461 ± 3
Relative latency	Fastest choice	Slower	Baseline
Reference fidelity	Improved	Improved, precision-focused	Strong
Multi-turn editing	Improved	Improved	Less optimized
Quality settings	Up to max	Up to max	Up to high
Custom dimensions	Yes	Yes	Yes
4K-class output	Yes	Yes	Yes
Transparent backgrounds	Yes	Yes	Yes
Best use	Apps, social, prototyping, batch generation	Campaigns, polished products, difficult edits	Existing integrations

If image quality is the only question, Sunburst leads right now. If you want to know which model is easiest to build a responsive product around, start with Flare.

One common assumption needs correcting: Flare doesn't have a lower list price than Sunburst. Both use identical official token rates. Flare's economic edge comes from generating faster, which cuts waiting time, infrastructure concurrency, abandoned requests and iteration time.

Budgeting

OpenAI's current rates, identical for both variants:

Token type	Flare	Sunburst
Text input	$5 / 1M tokens	$5 / 1M tokens
Cached text input	$1.25 / 1M tokens	$1.25 / 1M tokens
Image input	$8 / 1M tokens	$8 / 1M tokens
Cached image input	$2 / 1M tokens	$2 / 1M tokens
Image output	$30 / 1M tokens	$30 / 1M tokens

An editing request is billed from the separate token categories the response reports:

cost = (text tokens × $5 + image input tokens × $8 + image output tokens × $30) ÷ 1M

Any cached tokens are charged at the cheaper cached rates from the table. Reference-image workflows are best budgeted from the token counts that real requests return, since text input, image input, cached inputs and image output are all priced separately. If you reach the models through a third-party gateway, its prices may differ from the official ones and can change, so check its live catalog before deploying.

Where it fits

Product photography and e-commerce

A retailer can keep the same product while changing the background, lighting, seasonal setting, shadows, composition or promotional treatment. Reference fidelity counts for more here than an attractive but slightly different product.

Campaigns

Sunburst helps when one master creative needs several tightly controlled variations. The product and brand layout stay put, and the team varies the headline, language, scene, offer or visual style around them.

Social and high-volume content

Flare's speed makes it the natural engine for social image generators, creator apps, personalized thumbnails, memes, avatars and automated pipelines. Products like these benefit more from fast iteration and throughput than from squeezing out a few extra preference points.

Prototyping

Before a team invests in detailed design, either variant can sketch app screens, site concepts, packaging, interiors, product ideas, slide visuals and ad layouts.

Editing tools for end users

Better preservation matters once people upload their own pictures. Trying on outfits, replacing a backdrop, restyling a room, producing headshot variants, retouching a product or personalizing an ad all hinge on changing one thing and leaving the rest alone.

Transparent assets

Built-in transparency is handy for cutout objects, stickers, interface graphics, overlays and anything headed into a compositing pipeline.

Limits to plan for

• Text in images still needs review. Rendering has improved, but advertisements, packaging, menus, legal copy and UI screenshots shouldn't ship without a human check.
• A prompt isn't a layout engine. Telling the model to put the logo exactly 40 pixels from the top-left corner isn't the same as a constraint in a design tool. When pixels matter, generate the picture and add exact text or interface elements in code or a design app.
• Drift hasn't gone away. Multi-turn editing is improved but not flawless, so long chains, recurring characters and big brand campaigns still need checking.
• High resolution is experimental. The API allows 8,294,400 total pixels and an edge of up to 3840, but anything above 2560×1440 is experimental. Test the exact high-resolution settings your pipeline needs and don't assume every 4K-class output behaves the same.
• Hard requests can be slow. Build in loading states, timeouts, retries for transient errors and asynchronous processing for batch jobs.

Picking a variant

Flare fits:

• Image apps aimed at consumers
• Generation at high volume
• Tools for social content
• Quick prototypes
• Generating visual-search results
• Interactive editors
• Any product where users expect fast feedback

Sunburst fits:

• Campaign assets headed for production
• High-end product imagery
• Reference-image edits that leave little room for error
• Hard revisions that take several steps
• Final visuals where exact detail beats latency

For most new apps, Flare is the default and Sunburst is the escalation path, used when a request is unusually difficult or the output is important enough to justify the extra wait.

Summary of the release

Treat this as a workflow upgrade, not just prettier images. The biggest gains are in control: subjects and products stay recognizable, edits stay local, changes persist from turn to turn, detailed briefs get followed, and you get transparent output plus flexible high-resolution sizes. Flare is the broadly useful model, with better quality and much lower latency, and it suits interactive and high-volume apps. Sunburst is the precision option. Its Arena results lead both Flare and Image 2 in text-to-image and single-image editing, so it's the more attractive pick for polished creative work and hard revisions.

You needn't commit to one model forever, either. Send ordinary requests to Flare, route precision-sensitive work to Sunburst, and measure latency, token use, edit accuracy and accepted-output rate inside your own product.

FAQ

What is GPT-Image-2.5?

OpenAI's image generation and editing family from September 2026, offered in the API as Flare and Sunburst.

How is ChatGPT Images 2.5 different from GPT-Image-2.5?

The first is what you use inside ChatGPT. The second is how developers get the same generation, through the API model IDs gpt-image-2.5-flare or gpt-image-2.5-sunburst.

What is Flare?

The fast variant, and the one OpenAI recommends by default for most API apps. It stresses good image quality, editing ability and noticeably lower latency.

What is Sunburst?

The higher-capability variant for generation and editing, meant for workflows where precision and control matter more than response time.

Is it better than Image 2?

Going by current evidence, yes. Arena's September 9 snapshot gave Sunburst 1421 ± 13 and Flare 1399 ± 13 for text-to-image, with Image 2 at 1381 ± 4. In single-image editing the gap was wider.

Is it faster than Image 2?

Flare is. OpenAI says it offers higher quality than Image 2 with 50% lower latency. Sunburst trades speed for extra precision.

Can it make 4K images?

Yes. Both variants accept custom dimensions up to 8,294,400 total pixels and an edge up to 3840. Resolutions above 2560×1440 are currently experimental.

What does it cost?

Prices match across the two: $5 per million tokens of text in, $8 for image in, $30 for image out, with cheaper rates when inputs are cached.

Which variant should I use?

Flare for interactive, high-volume, consumer, social and prototyping work. Sunburst when final-image fidelity and precise editing outweigh latency.