Published September 10, 2026
OpenAI split its new image generation into two models, and both are now live on Secret Chat AI. GPT Image 2.5 Flare handles the quick, everyday work. GPT Image 2.5 Sunburst is there when precision matters more than the wait. They replace the single GPT Image 2 entry in our Image Generator, and Flare is now the model the generator opens on.
The surprising part is the price. OpenAI charges the two models the same rate per token. Sunburst is not a premium tier, a paid upgrade or a separate purchase. A careful Sunburst render can use more output tokens and therefore more credits, but it does not sit on a higher price list.
One Release, Two Models
OpenAI released GPT Image 2.5 on September 8, 2026 as two separate models: 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.
OpenAI calls Flare "our fastest model for high-quality, everyday image generation" and says it delivers higher-quality images than GPT Image 2 at roughly 50% lower latency. Those are OpenAI's claims; we haven't benchmarked either one ourselves. Sunburst is positioned for high-precision generation and editing, best "for workflows where editing precision matters most." OpenAI is equally plain about the tradeoff: it takes longer to generate.
Underneath that difference, the two have a lot in common. Both accept text and image input, return images only, support inpainting, and run on the same image-generation and image-edit endpoints. They also offer the same quality settings: low, medium, high, xhigh, max and auto. Secret Chat uses auto, the default, so the model decides how much work the request calls for.
The models support transparent backgrounds when PNG or WebP is requested. Secret Chat asks for JPEG, so images generated here come back as JPEG.
The Price List Is Identical
This deserves the numbers, because the names make it easy to assume Sunburst belongs to a dearer tier. OpenAI's API changelog entry for September 8 says both models "use GPT Image 2 token rates." OpenAI bills both models at $5 per 1 million text input tokens, $8 per 1 million image input tokens, $1.25 per 1 million cached text input tokens, $2 per 1 million cached image input tokens, and $30 per 1 million image output tokens. Those are also the rates carried by the model they replaced.
What you see in Secret Chat credits is the work the render actually used. If Sunburst spends more image output tokens to produce a more careful result, the total rises; there is no honest fixed multiplier we can give you. The rate itself remains exactly the same as Flare's.
What Changed in the Image Generator
The old GPT Image 2 entry, backed by gpt-image-2, is gone. In its place are GPT Image 2.5 Flare and GPT Image 2.5 Sunburst. The picker now contains seven models in this order:
- GPT Image 2.5 Flare
- GPT Image 2.5 Sunburst
- Nano Banana 2
- Nano Banana Pro
- Grok Imagine 2.0
- Seedream
- Qwen Image 2 Pro
Flare takes the opening slot from Grok Imagine 2.0, which had held it since August 12. Here is the bit existing users need to know: if you have picked one of the six models still in the list, nothing moves for you. Your choice is remembered. The only exception is GPT Image 2: because that model is gone, its saved choice moves to GPT Image 2.5 Flare, the closest replacement. The new default is read only for someone who has never chosen a model.
Both GPT Image 2.5 entries expose the same controls. You can request up to ten variants from one prompt in one request, choose an aspect ratio, and set the resolution to 1K or 2K. They open on 1K.
Why There Is No 4K Setting
There is deliberately no 4K option on either GPT Image 2.5 model. Image output is billed per token, so pixels scale the wait and the price together, and OpenAI's own guide now calls resolutions above 2560×1440 experimental. If you arrive with 4K saved from a Nano Banana model, the setting steps down to 2K rather than resetting to the smallest choice.
The matching ceiling also matters inside a chat. The router can choose which GPT Image model draws, while you may never see that decision. Giving one model a taller ceiling would let an invisible routing choice change the size of your picture. Both stop at 2K on purpose.
Generate, Attach, Then Keep Talking
For a new picture, choose the Image Generator, open Settings if you want Sunburst or a different size, and write the prompt. You can also attach an existing image to either GPT Image model and say what needs to change. A follow-up in the same chat carries the conversation forward, which means "make it darker" works without another full description.
Create a clean product photograph of a dark green fountain pen beside a cream notebook, with the words FIELD NOTES printed exactly on the cover, viewed from directly above in soft window light.
That prompt gives the model an exact-text requirement, a fixed layout and a product-image use case. Those are the sorts of constraints that make Sunburst the deliberate manual choice. For loose exploration or several first attempts, start with Flare and spend the saved time looking at actual results.
Smart Agentic ChatGPT Can Choose the Artist
The most interesting change isn't in the Image Generator picker. In ChatGPT's default Smart Agentic mode, the model already decides whether your message needs a quick answer, a web search, deeper research or a picture. Until this release, "a picture" led to one OpenAI image model. Now the router chooses between Flare and Sunburst. This is the first time a Secret Chat router has chosen between two image models at all.
The rule is intentionally one-sided. Flare draws by default. Sunburst enters when you ask for a better rendering of a picture already in the conversation: the image is blurry, a word is misspelled, you want it "in better quality," or one exact element needs changing. The router can also pick Sunburst when the first request states a precision requirement outright, such as exact text, an exact layout, or a print or product image.
Why wait for that signal? On the first attempt, nobody knows whether the fast model will be good enough. Sending every first draft to Sunburst would spend your time on a judgment you haven't made. Once you have seen the picture and said where it falls short, the judgment exists, and the extra care has a reason.
The manual Create Image tool always uses Flare. Choosing that tool means you chose an image, not a longer wait. You can learn more about the broader decision process on our private agentic AI models page.
Only ChatGPT routes between two image models this way, because only OpenAI offers the pair at the same per-token price. The Nano Banana models are billed per token too, but on different price lists: Nano Banana 2 costs $0.50 per 1 million input tokens and $60 per 1 million image output tokens, while Nano Banana Pro costs $2 and $120. Choosing between them is your spending decision, so it stays with you in Settings rather than with a router.
What Privacy Means for an Image Request
Secret Chat remains a gateway. Your prompt goes to OpenAI through us, so you don't need an OpenAI account and the request is not attached to your identity on OpenAI's side. Registration asks for an email address alone, with no name or phone number. Generated images are stored in your own browser, not in an archive on our servers.
Every attached image has its EXIF metadata stripped by default before it is sent. That includes GPS data, capture time and camera serial information, on every file, with no setting to hunt down. Our article on EXIF metadata in AI uploads explains what can hide in an ordinary photograph.
There are two limits worth saying out loud. A gateway anonymises who is asking, not what is asked: OpenAI still reads what you type so its model can answer, and the usual judgment about what belongs in a prompt still applies. Metadata stripping doesn't redact the picture either. A photograph still shows whatever it shows.
The private AI models page explains the gateway model across providers. If the job is text rather than an image, the AI Council can ask several providers the same question and show where their answers disagree; it does not generate or compare pictures.
Frequently Asked Questions
- What are GPT Image 2.5 Flare and Sunburst?
They are two separate OpenAI image models released on September 8, 2026. Flare is positioned for fast everyday generation; Sunburst is positioned for high-precision generation and editing and takes longer.
- Is GPT Image 2.5 Flare now the default on Secret Chat?
Yes, for someone who has never chosen an image model. If you picked one of the six models still in the list, Secret Chat remembers it and nothing changes for you. A saved GPT Image 2 choice moves to GPT Image 2.5 Flare because the older model is gone.
- Does Sunburst cost more than Flare?
OpenAI charges identical per-token rates for both. A Sunburst render may spend more image output tokens and therefore more credits, but Sunburst has no higher rate and is not a paid upgrade.
- When does Smart Agentic ChatGPT choose Sunburst?
It chooses Sunburst for a stated precision requirement, or when you want a better or more exact revision of a picture already in the conversation. Flare handles everything else by default.
- Which model does the manual Create Image tool use?
It always uses GPT Image 2.5 Flare. In the Image Generator itself, you can choose Flare, Sunburst or any of the other five models from Settings.
- Can GPT Image 2.5 edit an image I upload?
Yes. Both models accept an attached image and an instruction, and a later message in the same chat can refine the result without repeating the full description.
- Can GPT Image 2.5 generate 4K images on Secret Chat?
No. Both entries offer 1K and 2K and open on 1K. A saved 4K choice from a Nano Banana model steps down to 2K when you switch to either GPT Image model.
- Do I need an OpenAI account to generate an image?
No. Secret Chat sends the request through its gateway, so it isn't attached to your identity on OpenAI's side. You register for Secret Chat with an email address alone.