GPT-Image-2.5 Flare vs Sunburst: Which One a Marketing Team Should Be Paying For
On September 8, 2026, OpenAI released ChatGPT Images 2.5 and, for the first time, split its image API into two models: gpt-image-2.5-flare and gpt-image-2.5-sunburst. Most marketers will never call either directly. But most marketers now use at least one tool that generates images for them, and from today, every one of those tools has a choice to make on your behalf. This post is for the person who picks that tool, budgets for it, or is asked “which model does it run?” by someone who does.
The facts below are from OpenAI’s model pages and launch coverage as of September 8, 2026. Re-check them before you commit to a number; OpenAI adjusts pricing, quality tiers, and model availability without much ceremony.
Same rate card, two speeds, one real difference
Start with what is identical. Both models accept text and image input and produce images only. Both serve the same two endpoints: image generation and image edits, with inpainting supported. Both offer six quality settings: low, medium, high, xhigh, max, and auto. Both bill on the same schedule:
| Per 1M tokens | GPT-Image-2.5 Flare | GPT-Image-2.5 Sunburst |
|---|---|---|
| Text input | $5.00 | $5.00 |
| Text input, cached | $1.25 | $1.25 |
| Image input | $8.00 | $8.00 |
| Image input, cached | $2.00 | $2.00 |
| Image output | $30.00 | $30.00 |
Text output is not billed, because the model does not produce any. Rate limits are shared too, from 100K tokens and 5 images per minute at Tier 1 up to 8M tokens and 250 images per minute at Tier 5.
OpenAI’s pages also state that these token rates match GPT Image 2. You will find at least one launch-day summary claiming the 2.5 rates are exactly double the GPT Image 2 rates; OpenAI’s own model pages do not say that, and we are going with the source. What OpenAI does say, in a footnote worth reading twice, is that its GPT Image 2 cost calculator does not estimate 2.5 token consumption. Same price per token does not guarantee the same tokens per image.
So the difference is not price and not capability on paper. It is what each model was tuned to do:
- Flare is the default. OpenAI positions it as higher quality than GPT Image 2 at up to 50% lower latency, for “creator and social content to product experiences, visual search, rapid image prototyping, and high-volume generation.”
- Sunburst trades generation time for “tighter control across edits,” and OpenAI names the use cases directly: “production-ready campaign creative and polished product imagery.”
In OpenAI’s words to developers: “Choose Sunburst for workflows where editing precision matters most, and Flare for fast, high-quality everyday image generation.”
Where Sunburst earns its wait
Editing precision sounds abstract until you have lost an afternoon to its absence. The failure it addresses is the one every marketer who has generated a hero image knows: the first draft is 85% there, you ask for the one change, and the model helpfully changes four other things. Now the product label is different, the shadow direction flipped, and the version you liked is two turns back.
Sunburst is built for the images that will go through that loop several times. In marketing terms, that is a short list:
- The campaign hero. One image, used on the landing page, the announcement email, and the pinned post. It will be edited four to six times because it is the one image everyone has an opinion about.
- Product imagery. Anything where a real product from a reference photo has to survive a scene change with its proportions, label, and color intact.
- Anything with a named person or character. Founder illustrations, a brand mascot, a recurring character in a series. Drift across edits is most visible on faces.
For these, the extra generation time is cheap against the cost of the sixth regeneration.
Where Flare is plenty
Everything else, which is most of what a marketing team produces by count:
- Social tiles and LinkedIn illustrations, where the image is seen for two seconds and the set matters more than any single frame
- Blog covers and email headers
- First drafts of anything, including the hero, before the edit loop begins
- Concept exploration: five directions for one brief, pick one, refine that one (possibly on Sunburst)
Flare is also the model to run when volume is the point, because latency compounds. If an automation generates a header for every send in a sequence, or a tool produces a first image for every draft, halving the wait is the difference between a step you notice and a step you do not.
The rule is per job, not per team. The wrong way to read the split is “small teams get Flare, agencies get Sunburst.” The right way is that a single campaign uses both: Flare for the twelve supporting images, Sunburst for the one that carries the offer. A tool that lets you, or decides for you sensibly, is doing the job. A tool that runs everything on one model is either overpaying in time or underdelivering on the image that mattered.
What the quality tiers do to your bill
The six quality settings are where the real spend decisions live, and they are the part most marketing tools hide. Output is billed per image token, and higher quality tiers produce more tokens for the same pixel dimensions. OpenAI has not published per-image figures for 2.5, so we will not invent them. What we can say from running GPT Image 2 in production at Marqeable is that moving one tier from medium to high multiplied our cost per image by roughly five, with a visible improvement on print-sized heroes and no visible improvement on a 1080-pixel social tile viewed on a phone.
Two consequences:
autois a spending decision made by someone else. It lets the model pick the tier. That is convenient for a chat window and dangerous for a monthly invoice, because the model will reach for quality whenever the prompt sounds important.- The tier should follow the destination. Social and email at
mediumor below. Landing page heroes and anything that might be printed athighor above. If a tool cannot tell you which tier it uses for which asset, assume it uses one for everything.
The number that actually matters, though, is not price per generation at all. It is price per accepted image: every generation you paid for divided by every image that shipped. A model that costs 30% more per call and needs half as many calls to reach done is cheaper. That is the whole argument for Sunburst on edit-heavy work, and it is why What AI Image Generation Actually Costs a Marketing Team spends more time on regeneration rate than on rate cards.
How to read a vendor who says “we use GPT Image”
From today, that sentence is incomplete. A vendor that generates images for you now has four decisions baked into the product, and each one is a fair question:
| Question | What a good answer sounds like | What a weak answer sounds like |
|---|---|---|
| Which 2.5 model, for which job? | ”Flare by default, Sunburst for hero and product images” or a clear reason for one | ”The latest OpenAI model” |
| Which quality tier, and who controls it? | ”Pinned per asset type; you cannot accidentally run max" | "Auto” |
| How many images per request? | ”One, then you iterate" | "We generate four and you pick” |
| Do reference images go in? | ”Your brand board is attached to every generation" | "You can upload one in the prompt” |
The third row deserves a note. Generating four drafts so the user can pick one is a common pattern and it quadruples the bill for a preview most people glance at. Combined with auto quality, it is how a modest image feature becomes the largest line on an AI invoice.
What Marqeable runs
For the record, since we are asking vendors to answer: Marqeable generates images with OpenAI’s GPT Image 2.5 models, one image per turn, with the quality tier pinned in code per job so a prompt cannot escalate it, and with your brand profile’s palette, imagery style, exclusions, and reference images attached to every generation. The model is chosen per job the way this post describes, and we re-run the six tests from How to Choose an AI Image Generator for Marketing whenever OpenAI ships a new version. If you want to see it on your own brand, we are in private beta with a small early cohort: get early access.
Frequently asked questions
What is the difference between Flare and Sunburst?
Same pricing, same endpoints, same quality tiers. Flare is the fast default for everyday and high-volume generation. Sunburst is slower and tuned for control across successive edits, which OpenAI aims at campaign creative and product imagery.
Is GPT Image 2.5 more expensive than GPT Image 2?
Not per token, according to OpenAI’s model pages. Tokens consumed per image may differ, and OpenAI says its GPT Image 2 calculator does not apply to 2.5, so measure before you budget.
Can I use GPT Image 2.5 in ChatGPT, or only via the API?
Both. ChatGPT Images 2.5 rolled out to all ChatGPT tiers on the same day; the chat product does not expose the Flare/Sunburst choice or quality tiers. ChatGPT Images 2.5 for Marketers covers the chat side.
Should I switch a running workflow from GPT Image 2 to 2.5 today?
Test first. Run your last twenty prompts with their references through 2.5 and compare accepted-image rate and tokens per image against GPT Image 2. Switch when the numbers say so, not the announcement.
The bottom line
The split is simple once you stop reading it as a tier list. Flare is GPT Image 2 done faster and better, for the many images that just need to be good. Sunburst is for the few images that will be edited until they are right, and it earns its longer wait on exactly those. Price per token did not change; tokens per image might have, so measure. And from now on, “we use GPT Image” is a question, not an answer. Ask which one, for what, at which quality, how many at a time. The vendor who can answer has been paying attention to your images, and your invoice.
Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access
