GPT Image 2.5 vs GPT Image 2: Which to Use in 2026

TL;DR — OpenAI's 8 September 2026 release of GPT Image 2.5 replaces the single gpt-image-2 model with gpt-image-2.5-flare and gpt-image-2.5-sunburst, expanding quality settings from three to five. Standard token prices match GPT Image 2, but neither 2.5 model supports the Batch API. If you run bulk image jobs at half rate today, moving to 2.5 doubles your bill.

Verdict: Stay on GPT Image 2 for automated bulk jobs, and move to GPT Image 2.5 for live apps that need transparent cutouts or the new top quality tiers.

The upgrade to GPT Image 2.5 looks free when you look at the rate card. Standard token prices match down to the cent across both new options.

So your single-image test runs will look identical on your invoice. But the move isn't free for teams running batch jobs overnight. If you rely on OpenAI's batch pricing to cut your generation bills in half, GPT Image 2.5 breaks your budget.

What actually changed between GPT Image 2 and GPT Image 2.5

OpenAI replaced a single model with two distinct choices. You now pick between gpt-image-2.5-flare for fast responses and gpt-image-2.5-sunburst for detailed creative work.

OpenAI now files GPT Image 2 under a documentation section headed "Earlier GPT Image models". In that guide, OpenAI writes: "For new integrations, use one of the GPT Image 2.5 models described above." That's OpenAI telling you which way it expects developers to move.

The release also adds two higher quality steps and native transparent backgrounds. You can now request transparent PNG or WebP files right in your generation call without running a separate cutout tool.

Both new models ship with default snapshots dated 2026-09-08. Both carry the same standard token rates as the model released in April 2026.

GPT Image 2.5 vs GPT Image 2 at a glance

	GPT Image 2.5	GPT Image 2	What it means for you
Speed and the wait	Flare claims up to 50% lower latency; Sunburst takes longer on purpose	The baseline both claims are measured against	"2.5 is faster" is true of one of the two models, not of the release
Quality and detail	low, medium, high, xhigh, max, auto	low, medium, high, auto	Two more rungs at the top, and no published token count for what they cost
Editing and safety	1.09% unsafe shown on Sunburst, 1.41% on Flare	1.64% on the Images 2.0 baseline	Both beat the older baseline, but the card says the gap is not statistically significant
API models	Two: `gpt-image-2.5-flare`, `gpt-image-2.5-sunburst`	One: `gpt-image-2`	You now pick a lane before you write the call
Standard token rate	$8.00 in, $30.00 out per million	$8.00 in, $30.00 out per million	Identical, so the bill turns on tokens spent per image
Batch API	Not supported	Supported, at $4.00 in and $15.00 out per million	Moving a nightly job to 2.5 doubles its rate
Transparent background	Documented setting from launch	—	Cutouts without a second tool
Watermarking	C2PA plus SynthID	C2PA	Anything you generate carries an invisible mark as well as metadata

A dash (—) means OpenAI does not document the feature for that model. Sources: https://developers.openai.com/api/docs/pricing and https://deploymentsafety.openai.com/chatgpt-images-2-5 Checked 2026-09-09.

Speed and the wait

OpenAI claims Flare delivers higher-quality images than GPT‑Image‑2 at 50% lower latency. That speed gain belongs entirely to Flare.

Sunburst is built differently on purpose. OpenAI describes Sunburst as offering an extra level of precision for detailed creative work with longer generation times.

So calling GPT Image 2.5 faster as a single upgrade is simply wrong. One model gives you a shorter wait, while the other makes you wait longer to add precision.

Rate limits are identical across both 2.5 models. OpenAI caps usage by tier, measured in images per minute:

Tier 1: 5 images per minute.
Tier 2: 20 images per minute.
Tier 3: 50 images per minute.
Tier 4: 150 images per minute.
Tier 5: 250 images per minute.

Here's what those limits mean for your app. Picking Sunburst over Flare costs you nothing in volume caps, so your decision is purely about the wait per image.

Both models still share common generation limits. OpenAI notes that complex prompts may take up to 2 minutes to process across the board.

The same cold-brew bottle prompt rendered by GPT Image 2 on the left and GPT Image 2.5 Flare on the right, at the same size and aspect ratio

The same prompt, shape and size on GPT Image 2 and on Flare. On one product shot at this size, you can't see the upgrade — 2026-09-09.

Quality and detail

GPT Image 2 capped its output options at high. The new 2.5 models add xhigh and max quality settings alongside low, medium, high, and auto.

OpenAI claims 2.5 is better at preserving subjects from reference photos. It also claims the new release handles complex compositions and multi-element scenes with tighter adherence.

You can also set custom dimensions on 2.5. The width and height must be multiples of 16, aspect ratios must sit between 1:3 and 3:1, and total pixels must stay between 655,360 and 8,294,400. OpenAI flags sizes above 2560x1440 as experimental.

Here's the catch with image detail: OpenAI hasn't published token counts for 2.5. A rate card is not a price per image, because your total cost is the rate multiplied by the tokens spent.

We know standard rates match, but we don't know how many tokens xhigh or max consume per call. You can use our GPT Image cost calculator to track older baselines, but the new top tiers will increase token spend.

Streaming also adds to your invoice. If you stream a generation so users see partial output as it builds, each partial image costs an extra 100 image output tokens.

Both models still share familiar weaknesses. OpenAI documents that 2.5 can still struggle with precise text placement, recurring character consistency, and strict layout placement in structured compositions.

Editing and safety

OpenAI claims GPT Image 2.5 follows multi-turn editing instructions more reliably and modifies only the requested areas. Both 2.5 models support two endpoints: image generation and image edits.

Both models list exactly one supported editing feature: inpainting. Inpainting means you pick part of an image and describe what should replace it, leaving the untouched areas alone.

On safety, OpenAI published adversarial prompt tests in its system card. When deliberately attacked, Sunburst showed 1.09% unsafe images and Flare showed 1.41%, compared to 1.64% on the ChatGPT Images 2.0 baseline.

Those numbers come from a deliberately adversarial test set, not from regular traffic. Both 2.5 models beat the older baseline, but the system card states that no unsafe-shown difference reaches statistical significance at p < 0.05.

OpenAI also flags the new risk directly. The card notes that heightened realism could allow more convincing deepfakes of real people, places, or events if safeguards fail.

For provenance, 2.5 adds Google DeepMind's SynthID invisible watermarking alongside existing C2PA metadata. SynthID embeds an invisible watermarking layer across ChatGPT, Codex, and the API to mark generated content.

Independent developer Simon Willison reviewed the release on 8 September 2026 and noted on his blog that Sunburst is the stronger option based on the documented precision claims. He wrote: "Based on this I think Sunburst is the stronger option".

The batch discount is the part nobody mentions

The biggest practical difference isn't quality or speed. It's the fact that neither 2.5 model supports OpenAI's Batch API.

Under standard pricing, all three models cost $8.00 per million image input tokens and $30.00 per million image output tokens. Text input costs $5.00 per million tokens, and cached image input costs $2.00 per million.

But on OpenAI's batch pricing schedule, the GPT Image 2 Generator bills batch work at half the standard rate, $4.00 per million input tokens and $15.00 per million output. That gives high-volume teams an automatic 50% discount on non-urgent jobs.

Both gpt-image-2.5-flare and gpt-image-2.5-sunburst list the v1/batch endpoint as Not supported. Neither model appears anywhere in the batch pricing table.

Here's what that means for your production spend. If you run a nightly job generating thousands of catalog images on GPT Image 2, switching to GPT Image 2.5 will double your bill overnight.

Standard rates are identical, but without batch support, 2.5 forces you to pay full live rates for background work. For volume operations, that makes GPT Image 2 far more economical.

What OpenAI hasn't published yet

OpenAI has published no token counts for either 2.5 model, and says its GPT Image 2 calculator does not estimate them. So nobody outside OpenAI can yet say what a single image costs at xhigh or at max. Until someone measures it, the honest position is that 2.5 matches GPT Image 2 on the rate card and is unknown on the bill. We ran no benchmarks of our own.

Where the AI Overview is wrong about this comparison

Google's AI Overview summary makes three distinct errors about this comparison.

Speed & Latency: The Overview claims GPT Image 2.5 brings faster speeds as a general property. That is wrong because the 50% lower wait applies only to Flare, while Sunburst is designed to take longer for precision work.
Quality & Detail: The Overview presents 2.5 as superior across every axis. It ignores that 2.5 drops Batch API support, making GPT Image 2 half the price for bulk offline jobs.
Editing & Safety: The Overview lists safety as an outright improvement without context. While test numbers improved slightly on adversarial benchmarks, OpenAI's own system card confirms the differences are not statistically significant.

Run both generations on Seadanse today

You can run the original GPT Image 2 directly on Seadanse alongside tools like the Nano Banana 2 Generator, Seedream 5 Pro, Seedream 5 Lite, and Grok Imagine Image 2.

You can also run GPT Image 2.5 on Seadanse today through two separate options in the picker: GPT Image 2.5 Flare for fast everyday runs, and GPT Image 2.5 Sunburst for precise editing at longer generation times. Both take a prompt or up to three reference images at 1K, 2K, or 4K. On Seadanse, a 4K generation costs what a 1K draft costs on either model, though Sunburst costs more overall than Flare. The GPT Image 2.5 page has the full breakdown.

You can test models today with free starter credits on sign-up without entering a card. If you need more runs, check out Seadanse pricing for credit packs that don't lock you into a subscription, with automatic refunds on failed generations.

The Seadanse generator with GPT Image 2 selected as the image model

GPT Image 2 selected in the Seadanse generator — 2026-09-09.

When to move, and when to stay

Your decision comes down to how you call it and how long you can wait.

When to move to GPT Image 2.5

ChatGPT Interface: You use the web chat where batch discounts don't exist anyway.
Product Photos: You need native transparent backgrounds without running extra cutout software.
Rapid Prototyping: You can use Flare to get draft concepts back to users with a shorter wait.
Creative Work: You need the tighter control and top rungs offered by xhigh and max.

When to stay on GPT Image 2

Batch API: You run scheduled nightly jobs that rely on the 50% batch token discount.
High-Volume Budgets: You generate massive image sets where standard token rates would break your unit economics.
Stable Set-ups: You have something that works and don't need transparent output or experimental sizes.

GPT Image 2.5 vs GPT Image 2 FAQ

Which GPT model is best at image generation?

OpenAI calls gpt-image-2.5-sunburst its most capable image model, and points you to gpt-image-2.5-flare when you want the shorter wait. The two publish the same settings, the same rates and the same limits, so that line is positioning, not a spec.

How good is GPT Image 2?

GPT Image 2 remains a capable model that handles quality settings up to high and costs half as much when called through the Batch API.

Does GPT Image 2 still work?

Yes, gpt-image-2 remains fully supported in the OpenAI API under the earlier models section, with standard and batch endpoints active.

Is the GPT Image 2.5 upgrade free?

Standard token rates match GPT Image 2, but the upgrade isn't free if you run batch jobs, because 2.5 drops the 50% batch discount.

The verdict

GPT Image 2.5 is a capable upgrade for real-time apps, offering transparent backgrounds, a shorter wait on Flare, and two more quality settings than GPT Image 2 on both. But because both 2.5 variants drop Batch API support, GPT Image 2 remains the cheaper choice for high-volume offline jobs. Both generations sit in our picker right now, so you can test GPT Image 2.5 on Seadanse alongside GPT Image 2 to settle the upgrade question on your own prompt.

We read OpenAI's API guides, pricing pages and system card on 9 September 2026.

What Is GPT Image 2.5? — the full picture of what shipped on 8 September.
GPT Image 2.5 Flare vs Sunburst — which of the two to call, and why the docs won't tell you.
How to Use GPT Image 2.5 — the settings that matter, in ChatGPT and in the API.