GPT Image 2.5 Flare vs Sunburst: Which One to Use

OpenAI put ChatGPT Images 2.5 out on September 8, with two new models in the Images API: gpt-image-2.5-flare and gpt-image-2.5-sunburst. The token rates are identical to GPT Image 2, which sounds like a relief and isn't quite one. I'll come back to why.

The first decision is simpler than the pricing. One of these models is fast, the other is careful, and which one you call depends on whether anybody is going to stare at the image you just made.

Flare or Sunburst, in one table

Flare is the default and it deserves to be: better quality than GPT Image 2 at roughly half the latency. Sunburst is bigger and slower on purpose. The extra compute buys tighter control when you're editing the same image over and over, which is the difference between a model that helps and a model that undoes your last three instructions.

Start on Flare. Escalate only when the edit itself is the deliverable.

And gpt-image-2 is still listed, so nothing you've shipped is broken. You can take your time.

The two upgrades you'll actually feel

Subject preservation is the first one. A face, or some small identifying detail from a reference photo, now survives the generation instead of getting quietly reinvented into somebody's cousin. If your product builds variations on a customer's own upload, that has been the entire fight since version one.

Precision editing is the second. Ask for a new background and you get a new background, not a new composition with your product floating in it. Higgsfield AI's summary of this is the cleanest one going: the model understands what not to change.

Past those two, edits hold across a conversation, so the fifth instruction doesn't turn the picture to soup. Text renders sharper and transparent backgrounds come out cleaner, both of which Manus reported along with the speed figure. Light and surfaces look better too, which is the thing reviewers keep screenshotting.

C2PA metadata and the invisible watermark are on every output, same as before. Dull, and worth saying, because in some regulated workflows that's a requirement rather than a bonus.

Eight megapixels and two quality tiers above high

Both models sit on the generation and edit endpoints you're already calling. What's new is headroom.

Quality runs auto, low, medium, high, and then xhigh and max on top. Sizes break the old 2048×2048 ceiling and go to 3840×2160 landscape or 2160×3840 portrait, within a few rules worth writing down before you build a size picker:

• nothing over 3840px on an edge

• every dimension a multiple of 16

• aspect ratio no wider than 3:1

• total pixels between 655,360 and 8,294,400

That puts your biggest output just over 8 megapixels. 4K territory. Background takes auto, opaque or transparent, and that last value is how you ask for a transparent PNG.

Now hold those two facts together: outputs four times the old pixel count, and two quality tiers nobody has priced.

The price didn't change. That's the problem.

$8.00 per million image-input tokens, $30.00 per million image-output tokens, plus whatever your prompt costs in text. Same as GPT Image 2. For a reference point, Artificial Analysis measured GPT Image 2 at high quality at roughly $211 per thousand images, so about 21 cents for a 1024×1024.

Here's what an unchanged rate card hides. Per-image cost falls out of how many tokens a generation burns, and OpenAI didn't document that at launch. So the number you've been budgeting with, dollars per 1024×1024, is about to stop describing anything. Push to max at 8 megapixels and the cost per image goes somewhere. Nobody outside OpenAI can currently tell you where.

Sunburst will burn more than Flare on the same rate card, too. The gap between the siblings is real and unquantified, which is an awkward thing to write into a forecast.

Almost none of this has been checked independently

Every quality and speed claim above is vendor-reported. At launch neither model had an independent benchmark score, and no third party had reproduced the latency numbers.

OpenAI's own safety testing puts unsafe generations at 1.41% for Flare and 1.09% for Sunburst against a 1.64% baseline. That points somewhere useful without settling anything.

For what it's worth on adoption: Adobe is bringing both into Firefly, and Manus and Higgsfield AI are the other early customers OpenAI named.

The week this quietly costs you

A two-model split looks like a config change right up until you sit down to ship it.

Then it's token accounting you can't forecast. Retries when a generation times out. Polling for the slow model, webhooks, batching, and a fallback to gpt-image-2 while you work out which sibling has earned your traffic. If you've built that layer before you already know it isn't difficult work. It's just work, and none of it is your product.

Which is the case for an image layer that isn't welded to one vendor. APImage is the cleanest version of that I've used: one key, one base URL at https://app.apimage.org/api/v1, and a model field you change like any other string.

curl -X POST https://app.apimage.org/api/v1/image-studio \ -H "Authorization: Bearer sk_your_api_key" \ -H "Content-Type: application/json" \ -d '{ "prompt": "A minimalist Japanese garden at dawn", "model": "flux-2-pro", "aspect_ratio": "16:9" }'

Two things about it fit this particular mess. It bills in credits and reserves a job's cost up front, releasing it if the job fails, so a timeout doesn't charge you. That matters more than it sounds when you're poking at an undocumented cost curve on purpose. And the async side already exists: batch endpoints, webhooks, official JS/TS and Python SDKs, and long jobs that hand back an id you poll. Sunburst-length generation times go from a refactor back to a detail.

The model list covers both ends of the same workload. GPT Image 2 for text-heavy work like thumbnails and packaging, FLUX 2 Klein for fast iteration, Pro for everyday volume, Max when you need the hero shot. Up to 4K, transparent PNGs, four reference images per generation. Running one prompt through three models and keeping whichever came out right beats reading a launch page and guessing. There's an MCP server as well, twenty tools on one endpoint, if you'd rather point Claude or Cursor at it and skip the client code entirely.

Free accounts get three watermarked images, which is enough to find out whether the shape suits you. New models arrive on the same subscription and the same pool of credits, so a launch like this one costs you a string instead of a sprint.

Three numbers that'll settle it

Somebody outside OpenAI needs to reproduce Flare's latency claim. It's an afternoon's work for anyone with an API key, and Manus half-confirming it isn't the same thing.

The bigger unknown is token consumption per image, on both models, because until that's public the rate card doesn't convert into a price you can plan around.

And Sunburst has to place well on the image-editing leaderboards, because precision is the only reason it exists.

Until then: move to Flare and treat it as the straight swap for GPT Image 2. Hold out for Sunburst if your images get looked at one at a time by people who care. And build it so switching costs you a word, because there'll be another one of these launches before Christmas.

Note: APImage is not affiliated with OpenAI.