GPT Image 2.5 Guide: What Changed, Flare vs Sunburst, and Pricing
2026.09.10
"No more getting a different face every time you edit." OpenAI released GPT Image 2.5 on September 8 local time, and social feeds lit up. With the previous model, GPT Image 2.0, asking for a background swap would still shift a person's face a little. 2.5 targets exactly that problem. Features and prompts for 2.0 are collected in The complete ChatGPT Image 2.0 guide. This article covers the main changes, hands-on comparisons of Korean text rendering and Flare vs Sunburst, how to use it, pricing, and where it falls short.
What is GPT Image 2.5? (Flare and Sunburst)
GPT Image 2.5 is an image generation model OpenAI released on September 8, 2026, local time. Inside ChatGPT it looks like one model, but the API splits it into two: Flare and Sunburst.
What changed in GPT Image 2.5
There are four changes over the previous version. Here is what each one does to real work, with results.
1. The person stays the same across repeated edits
This is the change the video above shows. OpenAI says you can move a person or product from a reference photo into a new background and style. After the move, the person stays recognizable and their distinct features carry over. On lighting and texture, OpenAI's wording is that they look more natural, not that they are preserved.
2. Generation is up to 2x faster
OpenAI says it cut generation latency by up to 50% compared with Image 2.0. Some coverage cites "2 to 4 times faster," but the official announcement goes no further than "up to 50% lower latency."
3. It follows your hand drawings
Type
@Sketch in the ChatGPT composer to use a hand drawing as a reference. You set the composition and hand off the rest. The Carat editors fed in a ballpoint-pen scribble.
4. Edit by comment, export straight to a transparent background
Click a spot on a generated image, leave a comment, and only that part gets changed. The comment is an instruction, not a note. That is also where OpenAI's announcement stops listing 2.5 features. Templates for posters and merch arrive with it, so picking a template like "Poster" or "Merch" sets the layout. TechRadar's 24-hour hands-on lists the edit toolbar as Markup, Comment, Remove BG, Erase, and Resize. TechRadar also added a caveat that this toolbar appears to have rolled out quietly since August. Counting all five as new 2.5 features is a stretch.
You can get an image with a live alpha channel without erasing the background yourself. There is one catch. In the API, setting
background to transparent is not enough: you also have to set output_format to png or webp for the alpha to survive. The Carat editors missed this on the first test and got images with the background filled in white.🔥 What the Arena rankings actually show
Right after launch, Sunburst took first place and Flare second on the Arena leaderboard. What matters more than the ranking is which category gained, and by how much.
In text-to-image, Sunburst gained 40 points and Flare gained 18. Single-image editing was 59 and 30, multi-image editing 81 and 47. The gap widens as you move toward editing. The scores show plainly that this update is about editing, not generation. In absolute terms, text-to-image puts Sunburst at 1421, Flare at 1399, and GPT Image 2 at 1381, while Nano Banana Pro, a frequent point of comparison, sits at 1246 in 14th place.
How well does it write Korean
Korean text rendering is the first thing users check when picking an image model. The Carat editors ran the same prompt through Flare and Sunburst and checked the output directly.
You can run the prompt behind the image above on Carat as is.
👉 캐럿에서 지금 만들어보기
Large type was flawless on both models. Small type is where they split. In places where letters shrink, like the menu on a sidewalk sign, Flare smeared the strokes, while Sunburst kept most of it legible. Our team also tried work where a full Korean poster layout is involved.
👉 캐럿에서 지금 만들어보기
GPT Image 2.5 Flare vs Sunburst
Results from running both side by side on the same prompt. The Carat editors measured this on September 9, 2026.
Speed is the biggest split. On one image under the same conditions, Flare was about twice as fast as Sunburst. Arena scores go the other way, with Sunburst ahead. Text-to-image is 1421 to 1399, and multi-image editing is 1535 to 1501. If image quality comes first, use Sunburst. OpenAI's developer docs also describe Flare's quality as "similar to GPT Image 2."
How to use GPT Image 2.5
There are three main ways to use it. Here is how each one works, with the trade-offs.
1. Use it directly in ChatGPT
This is the fastest route. OpenAI says it is available on ChatGPT for desktop, mobile, and web, plus ChatGPT Work and Codex, on every plan. Open ChatGPT and ask for an image. Click the spot you want changed and leave a comment, or pick a template first for posters and merch to set the layout. One thing to know: OpenAI's launch announcement introduced Flare and Sunburst as models being added to the API, and the ChatGPT interface has no separate menu for choosing between them.
2. Use it on Carat
On Carat, Flare and Sunburst are listed separately, so you pick the one that fits the job. You can also set the quality tier yourself, from Low to Max. Right now, a new subscription or an upgrade to a higher plan comes with unlimited GPT Image 2.5. That condition can change over time, so check the terms that apply when you read this on the plan page. The interface is tuned for Korean prompts, so there is no need to translate to English: type it as is, for example "write 오늘의 커피 in Korean on a Seongsu-dong cafe sign." You can also generate with other models like Nano Banana Pro on the same prompt at the same time and compare.
3. Use it through the API (GPT Image 2.5)
The model names are
gpt-image-2.5-flare and gpt-image-2.5-sunburst. Quality has six tiers: auto, low, medium, high, xhigh, and max. GPT Image 2 topped out at high, and xhigh and max are new here. Prompts go up to 32,000 characters, up to 10 images per call, and up to 16 reference images. Output goes up to 3,840 pixels on the long edge.Prompt examples to copy and paste
The strength of 2.5 is in fixing rather than drawing from scratch. So how you write an edit instruction drives the result. Change what is inside the brackets and use it as is. If you would rather see results first and pick from there, GPT Image 2.5 prompt examples collects the outputs together with their prompts.
1. Change only what you want changed
This is the one line you will use most. It stops the 2.0 problem where "just change the shirt color" also changed the face.
Edit only the [part to change] that I specify. Do not touch the surrounding shape, material, lighting, or angle. Nothing else in the frame should change.
2. Drop your product into a new scene unchanged
Treat the uploaded product photo as fixed. Keep the shape, material, color, and logo exactly as they are. Place it in [scene] and adjust only the background, the lighting, and the camera angle.
3. Keep the layout, switch the text to Korean
Keep the product, the layout, and the lighting exactly as they are. Change only the visible copy to "[new copy]," and adjust background details minimally to suit the Korean market. Do not add any other text.
4. Cut it out onto a transparent background
Leave the product as is and remove only the background, placing it on a transparent background. Clean up only the ragged edges, and do not change the color, the shape, or the logo.
GPT Image 2.5 pricing
First, a common misunderstanding. Some coverage says the price doubled with 2.5. That is not true. Per OpenAI's official pricing page, Flare, Sunburst, and the previous GPT Image 2 all share the same token rates. Image input is $8 per million tokens, cached input $2, image output $30, and text input $5. GPT Image 2 has been on the same rates since July.
Per-image prices do not come out to round numbers because billing is by token. Even on the same model, resolution and quality tier change how many tokens are consumed, so OpenAI does not publish a list price per image and points to a calculator instead.
If you want to compare the cost against other image models, see the AI pricing comparison guide.
What are the generation limits
OpenAI did not publish per-plan generation limits with this release. The announcement only says "available on every plan." Figures going around, like "2 to 3 images a day on free" and "40 to 50 images per 3 hours on Plus," are community estimates from before 2.5.
Credits when using it on Carat
On Carat you pick the quality tier yourself, and each tier costs a different number of credits. Using Low as the baseline, Medium is about 2x, High about 5x, Extra High about 9x, and Max about 20x.
💬 Editor's note: The Carat editors ran the same prompt through every tier from Low to Max, and Korean accuracy was already perfect at Low. What improves as you go up is background detail. If you need text to come out right, Low is enough. For work where detail decides the result, like a close-up portrait or a product shot, go High or above.
🔥 The community split within a day
One day after launch, reactions did not settle on one side. Praise clustered around editing. Advertising practitioners pointed to earlier edits surviving repeated passes as the key gain, and one summary put it as image generation moving from magic to an editor. Criticism was about the look. The most widely shared posts compared it with Nano Banana Pro on the same reference, and the point was that the detail is overdone and looks artificial. There are also several reports of blotchy color in scenes with many elements.
GPT Image 2.5 pros and cons at a glance
✅ Pros
-
Edit consistency clearly improved. Arena multi-image editing gained 81 points over GPT Image 2, twice the gain in generation.
-
Korean rendering is stable, and you can pick the model that fits the job. Production assets like signs and posters came out accurate even at a low quality tier, and you can switch between Flare and Sunburst within the same pricing structure.
⚠️ Cons
-
Edits do not change only the area you asked about. OpenAI itself describes it as "better accuracy than before," not perfect preservation.
-
Opinions on the output style are divided, and there are reports of blotchy color in scenes with many elements.
-
Per-plan limits are not published, so there is no way to know how far the free plan goes.
Compare both models in one place on Carat
On Carat you can switch freely between ChatGPT Image 2.5 Flare and Sunburst, or generate with both on the same prompt and compare. You can also pick the quality tier yourself, from Low to Max.
👉 캐럿에서 지금 만들어보기
Frequently asked questions (FAQ)
Can I use ChatGPT Image 2.5 for free?
Yes, it works on the ChatGPT free plan. OpenAI says it is available on every plan across desktop, mobile, and web. Per-plan generation limits are not published, though, so there is no way to confirm how many images you get for free. On Carat, a new subscription or an upgrade to a higher plan comes with unlimited use. That benefit can change, so confirm the current terms on the plan details page before subscribing.
Should I use Flare or Sunburst?
It depends on the job. For speed and bulk generation, use Flare. For precise editing and top image quality, use Sunburst. OpenAI's docs suggest starting with Flare if the previous quality was good enough, and starting with Sunburst for complex work where it fell short. On Carat you can run both models on the same prompt and compare, so trying them yourself is the best way to decide.
Does Korean text really come out unbroken?
In the Carat editors' tests, large type like a shop sign or a poster title came out accurate from the Low tier on. In places where letters shrink, like a menu board, Flare sometimes smeared the strokes, so if small type matters, Sunburst is the safer choice.
Did the price go up from ChatGPT Image 2?
No, the token rates are the same. Per OpenAI's official pricing page, Flare, Sunburst, and ChatGPT Image 2 are all $30 per million tokens for image output. The batch discount that applied to ChatGPT Image 2 is not available on the two 2.5 models yet, so costs can differ for large batch jobs.
Can I use the images commercially?
Yes, commercial use is allowed within the scope of OpenAI's terms of use. Generated images do carry C2PA provenance metadata and an invisible watermark, and OpenAI describes these as a supporting measure rather than perfect detection. Cases involving brand logos or real people need separate review. See also What to check before using AI-generated images commercially.
War das hilfreich? Teile es mit deinen Freunden
Teilen
Täglich KI-Trends erhalten
