GPT Image 2.5 Flare vs Sunburst: What Actually Differs
GPT Image 2.5 Flare vs Sunburst specs are identical on paper. Here is what OpenAI actually published and how to pick the right one for your app.
The Seadanse Team11 min readThe team behind Seadanse. We run the models we write about, and we publish the specs, prices and limits the marketing leaves out.
TL;DR — OpenAI published identical specs for GPT Image 2.5 Flare and Sunburst, from rates to limits. The only documented differences are a single positioning line and a claim about generation speed. On 2026-09-09, neither model appears on independent leaderboards, so you can't rely on third-party test scores yet.
Verdict: Pick Flare as your default for speed, switch to Sunburst if you need precision editing, and test both on your own prompts.
You've got to drop a model id into your config file this week. You open the docs, expecting to see a clear technical tradeoff between Flare and Sunburst. But you won't find one there.
We put both documentation pages side by side on 2026-09-09 to see what separates them. The question is so new that Google search results don't even show an AI Overview for it yet. Here's what we found when we read every line.
What OpenAI says about each model
OpenAI launched both models on 8 September 2026. The official model page gives Flare a single summary line: "Fast, high-quality everyday image generation".
For Sunburst, the page says something else: "Our most capable model for image generation and editing".
The official image guide adds one more sentence of advice. The docs state to choose Sunburst for workflows where editing precision matters most, and Flare for fast, high-quality everyday image generation.
That is the only place in the reference docs where OpenAI separates the two models.
In the launch post, OpenAI calls Flare "the default choice for most applications". It says Flare delivers higher-quality images than GPT-Image-2 at 50% lower wait times.
OpenAI names five jobs for Flare:
- Creator and social content
- Product experiences
- Visual search
- Rapid image prototyping
- High-volume generation
OpenAI describes Sunburst differently in that same launch post. It says Sunburst "offers an extra level of precision for detailed creative work with longer generation times".
The company built Sunburst for premium creative jobs that benefit from tighter control across edits. It names two jobs for Sunburst: campaign creative and polished product imagery.
The model ids are gpt-image-2.5-flare and gpt-image-2.5-sunburst. Their default snapshots are gpt-image-2.5-flare-2026-09-08 and gpt-image-2.5-sunburst-2026-09-08. Aside from those names and the short blurbs, the two pages look like clones.
Simon Willison, creator of Datasette, looked at OpenAI's launch text on 8 September 2026. He wrote on his blog: "Based on this I think Sunburst is the stronger option".
He tested Sunburst that day on a reference image using a command-line tool he wrote himself.
Everything the two models publish identically
We checked OpenAI's pricing and guide pages on 2026-09-09. Every number, setting, and limit on Flare matches Sunburst line for line.
| GPT Image 2.5 Flare | GPT Image 2.5 Sunburst | What it means for you | |
|---|---|---|---|
| OpenAI's one-line pitch | Fast, high-quality everyday image generation | Our most capable model for image generation and editing | The only line on the two pages that differs |
| Model id | gpt-image-2.5-flare | gpt-image-2.5-sunburst | Switching is a one-string change |
| Image token rates | $8.00 in, $2.00 cached, $30.00 out per million | $8.00 in, $2.00 cached, $30.00 out per million | Same rate card, so cost turns on tokens spent per image |
| Quality settings | low, medium, high, xhigh, max, auto | low, medium, high, xhigh, max, auto | No quality ceiling either model can reach that the other can't |
| Supported endpoints | image generation, image edits | image generation, image edits | Same two calls |
| Supported features | inpainting | inpainting | Neither page lists anything the other lacks |
| Images per minute, Tier 1 to Tier 5 | 5, 20, 50, 150, 250 | 5, 20, 50, 150, 250 | Picking the slower model costs you nothing in rate limit |
| Generation time | OpenAI claims up to 50% lower than Images 2.0 | OpenAI says longer, on purpose | The one real difference, and OpenAI publishes no seconds for either |
| Independent Elo score | None on either Artificial Analysis board | None on either Artificial Analysis board | Nobody outside OpenAI has ranked them yet |
Every item on the reference sheet matches across both models:
- Token rates: $8.00 per million input image tokens, $2.00 cached, and $30.00 output image tokens.
- Text rates: $5.00 per million input text tokens and $1.25 when cached.
- Endpoints: image generation and image edits, with no other standalone endpoints supported.
- Features: exactly one supported feature listed, which is inpainting.
- Quality tiers: six options ranging from low to max, plus auto.
- Rate limits: Tier 1 gets 5 images per minute, up to 250 per minute at Tier 5.
The pricing page lists identical token rates for both models. Both pages say token rates match GPT Image 2, though the older calculator does not estimate 2.5 token use.
Note that while Responses isn't listed as a standalone endpoint, you can still use either model as the image tool inside the Responses API.
Both models share recommended sizes: 1024x1024 square, 1536x1024 landscape, and 1024x1536 portrait. Custom sizes must use multiples of 16, ratios between 1:3 and 3:1, total pixels between 655,360 and 8,294,400, and edges under 3840 pixels.
Both support transparent backgrounds, and both charge 100 extra image output tokens per partial image for streaming.
They share the same documented limits too. OpenAI notes complex prompts can take up to 2 minutes, and both models can struggle with text clarity, recurring characters, or structured layouts.
The one place OpenAI does separate them
The only clear gap OpenAI claims between Flare and Sunburst is generation time.
OpenAI states that Sunburst offers precision for detailed work with longer generation times on purpose.
Meanwhile, OpenAI claims Flare delivers images with 50% lower wait times than GPT Image 2.
But OpenAI publishes no speed numbers in seconds for either model. There are no wall-clock averages, no p50 numbers, and no p95 figures in the docs.
So this claimed gap is a positioning statement, not a verified measurement. We know Sunburst takes longer because OpenAI says it does.
But we don't know how many seconds you'll actually wait on any specific image size or quality setting.
Neither model has an independent score yet
When you can't find hard numbers in the docs, you check independent leaderboards. We checked Artificial Analysis on 2026-09-09.
On their text-to-image leaderboard, OpenAI's GPT Image 2 high ranks first with an Elo score of 1178 from 14,585 human comparison votes.
Elo here is a blind human-preference score. People pick between two generated images without knowing which model made each.
Google's Nano Banana 2 ranks fourth on that board at 1121. ByteDance's Seedream 5.0 Pro ranks thirteenth at 1083.
On their image-editing leaderboard, Microsoft's MAI-Image-2.6 leads at 1122. GPT Image 2 high ranks second at 1117.
You can test that older generation power today on our GPT Image 2 Generator.
Neither gpt-image-2.5-flare nor gpt-image-2.5-sunburst appears on either Artificial Analysis board as of 2026-09-09. We checked both the text-to-image and image-editing leaderboards.
Neither model has an Elo score, a win rate, or a public ranking anywhere.
So if someone tells you Sunburst scores higher on standard benchmarks today, they are guessing. Nobody outside OpenAI has published blind test data for these two models yet.
What the safety card shows, and what it doesn't
OpenAI published an adversarial safety evaluation in its system card.
On a test set of difficult prompts, Sunburst showed an unsafe image 1.09% of the time, Flare showed 1.41%, and the older ChatGPT Images 2.0 baseline showed 1.64%.
Lower numbers mean fewer unsafe images got shown. But you shouldn't use these numbers to rank the two models against each other.
The system card states that no unsafe-shown difference meets the threshold at p < 0.05 using a two-sided exact McNemar test.
The card says the 2.5 series generally performs on par with or better than 2.0. It notes that minor regressions do not meet the test threshold.
So the two look different on that table. But the card says the difference doesn't clear its own bar for significance, which means the numbers aren't far enough apart to tell the models apart. Treat them as tied on safety.
Remember also that this test set was deliberately adversarial. It tested how the models handle harmful inputs, not how they handle everyday prompts.
How to actually choose between Flare and Sunburst
Since the published specs are identical, you can't pick a winner from the spec sheet alone.
You have to test both models on your own images and prompts.
Set up a side-by-side test in your app. Send the exact same prompt, aspect ratio, and quality setting to both model ids.
Check these three specific things to see which model fits your job:
- The wait: time each generation with a wall-clock timer, and see if Flare finishes much faster on your typical image sizes.
- Editing precision: check whether an inpainting edit changes only what you asked for while leaving untouched regions clean.
- Consistency across edits: run a face or product image through several rounds to see if Sunburst preserves fine marks better.
The same prompt, shape and size on Flare and on Sunburst. Both get the label right. They differ in styling, not accuracy — 2026-09-09. If you don't need heavy multi-step edits, pick Flare. It's built as OpenAI's everyday default.
Try both against the models you can run today
GPT Image 2.5 runs on Seadanse today as two separate entries in the image model picker: GPT Image 2.5 Flare and GPT Image 2.5 Sunburst. Both take a prompt or up to three reference images, and both offer 1K, 2K and 4K outputs.
This is where your test takes one browser tab. Both models sit in the same picker under those exact names, so you can test the same prompt on each without an API key or writing any code. Sunburst costs more than Flare, and neither one costs more at 4K than at 1K.
You can also run the same prompt on our GPT Image 2 Generator to see how the older leader compares. Or test Google's Nano Banana 2 Generator and ByteDance's Seedream 5 Pro Generator on that same prompt.
We also host Seedream 5 Lite and Grok Imagine Image 2.
New accounts get free starter credits on sign-up without entering a card.
If you want more runs, you can buy a credit pack with no monthly subscription. Check our Seadanse pricing to see every option.
If any image generation job fails or gets interrupted, our system refunds your credits automatically.
The image model picker in the Seadanse generator. GPT Image 2.5 is two entries here, Flare and Sunburst — 2026-09-09.
Flare vs Sunburst FAQ
Which model is better: Flare or Sunburst?
Neither model is universally better. OpenAI positions Sunburst for precision editing and complex creative jobs, while Flare is the default for faster everyday image generation. Because independent benchmarks haven't ranked them yet, you must test both on your own prompts to decide.
Is Sunburst slower than Flare?
Yes, according to OpenAI. OpenAI states that Sunburst uses longer generation times on purpose to deliver tighter control across edits. But OpenAI publishes no official speed numbers in seconds, so you should time both models on your typical image sizes.
Do Flare and Sunburst cost the same?
Yes, their published token rates are identical. Both cost $8.00 per million input image tokens, $2.00 per million cached input tokens, and $30.00 per million output tokens. Your total cost will depend purely on how many tokens each model consumes to produce your image.
Can you switch between Flare and Sunburst later?
Yes, switching is as simple as changing a single string in your code. Because both models share the exact same quality settings, size limits, endpoints, and rate limits, swapping gpt-image-2.5-flare for gpt-image-2.5-sunburst needs no other code changes. If you use a browser generator that lists both by name, switching is as quick as picking the other entry.
The verdict
OpenAI gave two different names to models that share identical published documentation.
Their pricing, quality settings, endpoints, size limits, and rate limits match across every single row.
The only documented difference is OpenAI's positioning line: Flare is the fast everyday default, and Sunburst is the slower precision editor.
We recommend starting with Flare in your configuration file.
If your multi-step edits lose detail or drift, swap the model id to Sunburst and compare the results on the GPT Image 2.5 page.
We read and compared OpenAI's official model guides, pricing tables, and safety cards side by side on 2026-09-09 to write this report.
Related articles
- What Is GPT Image 2.5? — everything OpenAI shipped on 8 September.
- GPT Image 2.5 vs GPT Image 2 — the upgrade question, and the batch discount you lose.
- How to Use GPT Image 2.5 — the settings that matter, in ChatGPT and in the API.
