Why are GPT Image 2.5 Flare and Sunburst the same price?

OpenAI has not explained it, and the reasoning is worth thinking about because it changes how you should choose. The two models optimise for opposite things — speed against precision — so a rational price gap would push everyone to the cheaper one. Pricing them identically removes that pressure and lets the choice be made on the work rather than the budget. The practical consequence for you is that there is no cost reason not to default to Sunburst, and no cost penalty for switching per task.

What are the known limitations of GPT Image 2.5?

The main one is a migration trap rather than a model flaw: prompts written for GPT Image 2, and especially AI-generated reference photographs, have been reported to perform worse on 2.5 and need rewriting. Practitioners who dropped the generated reference image and fixed the prompt reported a clear improvement in likeness and in rendering text on surfaced objects. Its text rendering is better than the previous generation but is still the class-wide weak point, so check fine typography before it ships.

By Viral AI Trends · Published 7 October 2026

GPT Image 2.5: Flare or Sunburst, and what an image actually costs

Q: What is the difference between GPT Image 2.5 Flare and Sunburst?

Flare is the fast one and Sunburst is the precise one, and OpenAI describes them as different training objectives rather than different sizes of the same model. Flare carries a Small size tag and is positioned for social content, e-commerce imagery, visual search, rapid prototyping and high-volume generation, with throughput described as two to four times faster. Sunburst carries a Base size tag and is positioned for ad creative, high-end product shots and complex editing workflows, where generation takes longer in exchange for tighter control. Notably, Sunburst is also the model people report using for clean transparent-background cutouts.

Q: How much does GPT Image 2.5 cost?

Flare and Sunburst share identical token pricing: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens. At 1024 by 1024 that works out to roughly $0.006 at low quality through to about $0.211 at maximum quality. The pricing is unchanged from GPT Image 2 — but the high-quality tier now consumes the token budget that used to be the medium setting, which makes high-quality output roughly four times cheaper than the equivalent in the previous generation.

Q: Is GPT Image 2.5 free?

In ChatGPT, yes — GPT Image 2.5 is available across all tiers including the free one, so the model itself is reachable at no cost. The API is paid: there is no free tier and every image is billed. This is the sharpest practical difference between it and Google's Nano Banana 2.1, which is cheaper per image at the API level but has no equivalent free consumer route through a general assistant.

Q: Does GPT Image 2.5 support transparent backgrounds?

Yes, and it is the least-discussed capability of the release. Both models output true alpha-channel transparency in PNG, which means cutout work that previously meant a separate matting pass — e-commerce product images, game assets, UI icons, sticker packs — can now be generated with the background already gone. Practitioners have specifically credited Sunburst for producing clean transparent assets, which makes it the model to try first for that use case.

Q: Is GPT Image 2.5 backward compatible with GPT Image 2?

Yes. The endpoints are unchanged, so migration is a model identifier change: swap gpt-image-2 for gpt-image-2.5-flare for most cases, or gpt-image-2.5-sunburst where precision matters. GPT Image 2 remains available, so nothing breaks if you do not migrate — but you keep paying the old high-quality token budget if you stay.

Q: What resolution does GPT Image 2.5 support?

Up to 3840 by 2160, which is 4K, on both models. It accepts custom sizes in multiples of 16 and aspect ratios from 1:3 to 3:1, and it outputs PNG, JPEG or WebP. The combination of a 4K ceiling, arbitrary aspect ratios and alpha-channel output is what separates it from most consumer image tools, which force fixed shapes and flatten transparency.

Q: Is GPT Image 2.5 better than Nano Banana 2.1?

They are priced for different shapes of work and they are not interchangeable. Nano Banana 2.1 costs 3.36 cents at 1K on Google's published table against roughly 6 to 211 cents for GPT Image 2.5 depending on quality tier, so at volume Google is dramatically cheaper. GPT Image 2.5 gives you a free route through ChatGPT on every tier including Free, true alpha-channel transparency, arbitrary aspect ratios up to 3:1 and text-to-image prompts that fit its own vocabulary. Choose on which constraint binds: per-image cost favours Google, free access and cutout work favour OpenAI.

OpenAI released GPT Image 2.5 on 8 September 2026 and did something unusual with it: instead of one model it shipped two, tuned for opposite things, at exactly the same price. That removes the budget as a way to choose, which means you have to choose on the work. This page covers what separates Flare from Sunburst, what an image costs at each quality tier, why the high setting got roughly four times cheaper, and the capability almost nobody mentions — true transparent backgrounds.

What GPT Image 2.5 is

GPT Image 2.5 is OpenAI's image generation and editing model, released on 8 September 2026 as the successor to GPT Image 2. It does three things the previous generation did not, and one of them is not a quality feature at all.

Two API models rather than one. gpt-image-2.5-flare and gpt-image-2.5-sunburst. OpenAI describes them as separate training objectives aimed at speed and at precision respectively — not a large and a small version of the same model.
A 4K ceiling with arbitrary aspect ratios. Output goes to 3840×2160 on both models, sizes are accepted in multiples of 16, and aspect ratios run from 1:3 to 3:1. Output formats are PNG, JPEG and WebP.
Native alpha-channel transparency. Both models produce genuine transparent backgrounds in PNG rather than a checkerboard pattern, which is the feature with the largest practical consequence and the smallest amount of coverage.
Dual content provenance. Output carries both C2PA metadata and Google DeepMind's SynthID invisible watermark — an unusual pairing, and one aimed squarely at enterprise compliance review.

On the name. You will see this model written several ways, including "GPT Sunburst 2.5" and "GPT Image Flare". Those are not separate products — they are people spelling the API model identifiers gpt-image-2.5-sunburst and gpt-image-2.5-flare the way they read them. There is one model family here with two variants, and ChatGPT calls the whole thing Images 2.5.

Flare vs Sunburst — the choice OpenAI refused to make for you

Most vendors ship a fast model and a good model and price them a tier apart, which quietly makes the decision for the buyer. OpenAI did not do that here. Flare and Sunburst cost exactly the same, so the only thing left to choose on is the job.

	GPT Image 2.5 Flare	GPT Image 2.5 Sunburst
Model ID	`gpt-image-2.5-flare`	`gpt-image-2.5-sunburst`
Size tag	Small	Base
Built for	Speed and throughput	Precision and editing control
Reported speed	Two to four times the throughput	Slower generation
Positioned for	Social content, e-commerce imagery, visual search, rapid prototyping, batch generation	Ad creative, high-end product shots, complex multi-round editing
Price	Identical	Identical

Because the price is the same, the useful question is not "which is cheaper" but "which fails less for this task". Flare's advantage is throughput, so it pays off when the bottleneck is how many variations you can generate — A/B testing creative, filling a catalogue, prototyping a layout you will discard. Sunburst's advantage is control, so it pays off when the bottleneck is how many tries it takes to get one image right.

There is a third consideration the marketing does not mention: Sunburst is the one practitioners single out for clean transparent-background cutouts. If alpha output is why you are here, start with Sunburst rather than Flare, regardless of how fast your pipeline needs to be.

What it costs, and the four-times-cheaper detail

Token pricing is identical across both models and unchanged from the previous generation:

Item	Rate
Text input	$5 per 1M tokens
Image input	$8 per 1M tokens
Image output	$30 per 1M tokens

At 1024×1024 that works out from roughly $0.006 at low quality to about $0.211 at maximum quality. The headline rate did not move — and that is exactly why the interesting change is easy to miss.

The high-quality tier now consumes the token budget that used to be the medium setting. In plain terms: max-quality output in GPT Image 2.5 costs roughly a quarter of what the equivalent setting cost in GPT Image 2, for comparable or better results. Nobody announced a price cut, because on paper there wasn't one; what changed is how many tokens the top setting burns to reach the same place. If you had ruled out the high tier on cost, that assumption is now out of date.

How to budget it. Ignore the per-token rate and count the images. A pipeline generating fifty product variants a day at medium quality is a different number from one generating six hero shots at maximum, and the per-image spread is roughly 35× across the quality range. Set the quality tier per job rather than globally — that single decision matters more than any rate comparison.

The free route — and how it compares with Google

GPT Image 2.5 is available in the OpenAI API and in ChatGPT across all tiers, including Free. That is the sharpest practical distinction between this model and Google's Nano Banana 2.1, which is cheaper per image at the API level but has no equivalent route through a general assistant on a free account.

So the two compare like this: if your constraint is per-image cost at volume and you can write code, Google is dramatically cheaper — 3.36¢ at 1K against a range that starts around 0.6¢ and climbs past 20¢ for OpenAI, depending entirely on the quality tier you pick. If your constraint is not having a billing relationship at all, or needing a free way to make a handful of images, OpenAI is the one you can actually use today.

The consumer features that arrived alongside it matter for the same reason. ChatGPT gained Sketch, where you type @Sketch and draw rather than describe; comment-based editing, where you place a note on a specific part of an image — "make this darker", "change this to blue" — instead of re-describing the whole scene; and templates for posters, merchandise, flyers and product photos. None of those are API features. They are why the free tier is worth having even if you also pay for an API.

Transparent backgrounds, and why this is the sleeper feature

Both models output true alpha-channel transparency in PNG. That sounds minor and is not: background removal has been a separate paid step in almost every production pipeline since AI images became usable, and it is the step that most often ruins an otherwise good asset — matting eats edges, halos survive, semi-transparent details like hair and glass come back wrong.

Generating with the background already absent removes that step rather than automating it. The use cases are exactly the ones where cutout cost used to dominate the decision:

E-commerce product images — clean product on a white or transparent field, generated rather than photographed and masked.
Game and app assets — icons, items, UI elements that need to sit on an unknown background.
Sticker packs and overlays — the format lives or dies on edge quality.
Composite work — generating a foreground element to place over a scene you already have.

If you tried transparent output on an earlier generation and gave up, this is worth retesting — but test on your hardest case rather than a simple one, because a clean cutout of a hard-edged object is not evidence that a cutout of hair will hold up.

Migrating from GPT Image 2

The migration is deliberately boring, which is a feature. Endpoints are unchanged, and the change is the model identifier:

gpt-image-2 → gpt-image-2.5-flare for most use cases.
gpt-image-2 → gpt-image-2.5-sunburst where precision or transparent output matters.

GPT Image 2 remains available, so nothing breaks if you do not migrate — you simply keep paying the old, larger token budget at the high setting. Two things are worth doing before you switch in production, and both are cheap:

Re-test your existing prompts rather than assuming.
Prompts written for GPT Image 2 do not always carry over cleanly. The reported pattern is specific and worth knowing: prompts tuned for the older model, and particularly ones that supply an AI-generated reference photograph, have performed worse on 2.5. Where that happened, rewriting the prompt and dropping the generated reference improved both likeness and text rendering on surfaced objects. Budget an afternoon for prompt review, not for a find-and-replace.
Pick the quality tier per job.
The per-image spread is roughly 35× from low to max. If your code sets one global quality value, you are almost certainly overpaying somewhere and underdelivering somewhere else. Split it: low or medium for variants you intend to discard, maximum for the image you keep.

Getting better images out of it

Default to Sunburst, switch to Flare for volume.
Since the price is identical, there is no cost reason to avoid the precision model. Use Sunburst by default and move a task to Flare when the number of images, not their quality, is the constraint.
Rewrite prompts, do not port them.
The single most common cause of disappointment with this release is carrying an old prompt across unchanged. Say what the image should contain, drop generated reference photos, and let the model do the describing it is better at than the previous generation was.
Ask for transparency explicitly, in the format you need.
Request PNG output and state the transparent background in the prompt. The model can do it, but it is generating to your instruction — being vague about the background is how you get a nice picture with a white field instead.
Use comment-based editing before regenerating.
In ChatGPT, place the note on the part that is wrong rather than re-describing the whole image. You keep everything that already worked and pay for one edit instead of a fresh generation — the same principle as conversational editing in video models.
Check typography at full size, every time.
Text rendering improved generation over generation and is still the class-wide weak point. Fine details are where it fails, and failures are invisible in a thumbnail. Zoom to 100 percent on any surface that carries words before it leaves your machine.
Know what the provenance layer means for your channel.
Output carries both C2PA metadata and a SynthID watermark. That combination is an advantage in enterprise review and a constraint in publishing — check the platform's AI-disclosure rules before you ship, and do not assume either mark can be stripped.

Frequently asked questions

What is GPT Image 2.5?

OpenAI's image generation and editing model, released 8 September 2026 as the successor to GPT Image 2. It ships as two API models — gpt-image-2.5-flare for speed and gpt-image-2.5-sunburst for precision — both outputting up to 4K with native transparent backgrounds at identical prices. It runs in the OpenAI API and in ChatGPT on every tier including Free.

What is the difference between GPT Image 2.5 Flare and Sunburst?

Flare is the fast one, carrying a Small size tag and positioned for social content, e-commerce imagery, visual search, prototyping and batch generation, with throughput reported at two to four times higher. Sunburst carries a Base tag and is positioned for ad creative, high-end product shots and complex editing, where slower generation buys tighter control. Sunburst is also the variant people specifically credit for clean transparent-background cutouts.

How much does GPT Image 2.5 cost?

Both models share the same rates: $5 per million text input tokens, $8 per million image input tokens, $30 per million image output tokens. At 1024×1024 that is roughly $0.006 at low quality up to about $0.211 at maximum. The rates are unchanged from GPT Image 2 — but the high-quality tier now consumes the token budget that used to be the medium setting, making max-quality output roughly four times cheaper than the equivalent in the previous generation.

Is GPT Image 2.5 free?

In ChatGPT, yes — it is available across all tiers including Free, so the model is reachable at no cost. The API is paid with no free tier. That is the clearest practical difference from Google's Nano Banana 2.1, which is much cheaper per image via API but has no comparable free consumer route.

Does GPT Image 2.5 support transparent backgrounds?

Yes — true alpha-channel transparency in PNG, on both models. It removes a separate matting pass from workflows that lived or died on edge quality: e-commerce product shots, game and app assets, sticker packs, and foreground elements for composites. Practitioners credit Sunburst specifically for this, so start there if cutouts are your reason for using it.

Is GPT Image 2.5 backward compatible with GPT Image 2?

Yes. Endpoints are unchanged, so migration is a model identifier swap: gpt-image-2 to gpt-image-2.5-flare for most cases, or gpt-image-2.5-sunburst where precision matters. GPT Image 2 is still available, so nothing breaks if you wait — you just keep paying the older, larger high-quality token budget.

What resolution does GPT Image 2.5 support?

Up to 3840×2160 (4K) on both models, with custom sizes in multiples of 16, aspect ratios from 1:3 to 3:1, and PNG, JPEG or WebP output. The 4K ceiling combined with arbitrary ratios and alpha-channel output is what separates it from consumer image tools that fix the shape and flatten transparency.

Why are Flare and Sunburst the same price?

OpenAI has not explained it, and the reasoning matters because it changes how you choose. The two optimise for opposite things, so a price gap would steer everyone toward the cheaper one. Identical pricing removes that pressure and leaves the work as the deciding factor. The practical upshot: there is no cost reason not to default to Sunburst, and no cost penalty for switching per task.

What are the limitations of GPT Image 2.5?

The biggest trap is migration rather than the model: prompts written for GPT Image 2, and especially those supplying AI-generated reference photographs, have been reported to perform worse on 2.5 and need rewriting — after which likeness and text rendering improved. Text rendering is better than the previous generation but remains the class-wide weak point, so check fine typography at full size before shipping.

Is GPT Image 2.5 better than Nano Banana 2.1?

They suit different constraints. Nano Banana 2.1 is 3.36¢ per 1K image on Google's table, against a GPT Image 2.5 range of roughly 0.6¢ to 21¢ depending on quality tier — so at volume Google is usually far cheaper. GPT Image 2.5 offers free access through ChatGPT on every tier including Free, true alpha-channel transparency, aspect ratios up to 3:1 and a prompt vocabulary of its own. Choose on which constraint binds for you rather than on which is "better".

Also on this site

Nano Banana 2.1 — the direct competitor, released a month later and priced at roughly a quarter per image. What it wins on, and where Pro still beats it.
Gemini Omni — OpenAI's image competitor has a Google counterpart on the video side: the model Google built instead of a Veo 4, its free route and its real cost.
Seedance 2.5 — ByteDance's video model, and the four specifications no engineering source confirms.
MiniMax H3 — the video-side comparison: ranked first on Arena's image-to-video board, and one of the few frontier models with downloadable weights.
Hotel Lobby AI — the two-photo duet format that took over feeds in 2026: the walkthrough, the prompts and the permission rules.

Disclosure. We are an independent guide and not affiliated with OpenAI, Adobe, Manus or any platform listed. Prices, model identifiers and capabilities on this page were read from OpenAI's launch material and pricing, plus practitioner reporting that is named where used. Per-image figures are conversions from published token rates and move with resolution, quality tier and prompt length. Verify before purchase. Where a claim could not be confirmed against a primary source, the page says so. Some links may be affiliate links; they do not change what we recommend or how we rank anything.