What Is GPT Image 2.5? Two Models, Five Quality Tiers
GPT Image 2.5 brings Flare and Sunburst, five quality tiers, and shorter wait times to OpenAI's image API. Here is what changed and what it costs.
The Seadanse Team12 min readThe team behind Seadanse. We run the models we write about, and we publish the specs, prices and limits the marketing leaves out.
TL;DR — GPT Image 2.5 is OpenAI's image model released on 8 September 2026, offering two API models named gpt-image-2.5-flare and gpt-image-2.5-sunburst with five quality settings where GPT Image 2 had three. The update shortens the wait, adds transparent background support, and gives you two higher output tiers above high.
Need careful edits over several rounds? Test Sunburst on our GPT Image 2.5 Generator. Calling OpenAI's Batch API for the 50% discount? Stay on GPT Image 2, though batch generation is coming to Seadanse.
You probably read yesterday that OpenAI released a faster, sharper image system. Almost every article reporting on the launch repeated OpenAI's press release word for word.
None of them said what it costs, which of the two models to call, or what the new settings do to your bill. We read OpenAI's rate card, model pages, and system card on the day they went live to find the real numbers.
What GPT Image 2.5 changes in ChatGPT
ChatGPT Images runs on the new 2.5 system across web, mobile apps, and Codex. OpenAI reports that users create more than 3 billion images each week across ChatGPT and its image API.
The ChatGPT Images panel and its templates, captured 2026-09-09. The update focuses on five clear areas inside the ChatGPT interface:
-
Faster generation: OpenAI claims it reduced image generation latency by up to 50% compared with Images 2.0, speeding up conversational creation in standard chat threads. In plain terms, you wait less time for a picture.
-
Better detail: OpenAI says you get richer textures and more natural lighting, and that the model follows complex instructions more closely.
-
Precise edits: OpenAI says it changes only what you asked for and holds the rest of the picture still, even over several rounds of edits.
-
Sketch tool: On the phone app, type
@, pick Sketch, draw, then say what it should become. -
Templates & Sharing: You can open the Images tab in the sidebar to pick pre-built style templates, or share the exact prompt behind any creation by copying a prompt link. Templates aren't available in Work mode yet.
Every generated file includes C2PA metadata, and OpenAI now embeds invisible SynthID watermarking that complements C2PA metadata-based approaches across ChatGPT, Codex, and the API.
Run OpenAI's image models on Seadanse
Seadanse runs image and video models in one browser tab, so you can test two of them on the same prompt without signing up twice. Today, you can run GPT Image 2, Nano Banana 2 Generator, Seedream 5 Pro Generator, Seedream 5 Lite, and Grok Imagine Image 2 right in your browser.
You can run GPT Image 2.5 on Seadanse today as two separate options: GPT Image 2.5 Flare for fast everyday work, and GPT Image 2.5 Sunburst for precise editing at longer generation times. Both take a prompt or up to three reference pictures, and both offer 1K, 2K, and 4K output. Sunburst costs more than Flare, but a 4K render costs the same as a 1K draft on either model.
New accounts get free starter credits upon sign-up with no credit card needed. There's a credit pack that doesn't need a subscription.
If a generation fails or gets interrupted, the credits go back automatically. You can review all account options on our Seadanse pricing page.
The image model picker in the Seadanse generator. GPT Image 2.5 is two entries here, Flare and Sunburst — 2026-09-09.
Flare and Sunburst are the same model on paper
OpenAI split the 2.5 release into two API model identifiers: gpt-image-2.5-flare and gpt-image-2.5-sunburst. Default snapshots point to gpt-image-2.5-flare-2026-09-08 and gpt-image-2.5-sunburst-2026-09-08.
OpenAI positions Flare as its everyday engine for rapid prototyping, social media content, and high-volume product images. It positions Sunburst as the premium option for demanding commercial creative work where fine control across successive edits matters most. But when you inspect the published numbers on OpenAI's documentation pages, both models look identical.
| GPT Image 2.5 Flare | GPT Image 2.5 Sunburst | What it means for you | |
|---|---|---|---|
| OpenAI's positioning | Fast, high-quality everyday image generation | Our most capable model for image generation and editing | The only difference OpenAI puts in writing is this row |
| Model id | gpt-image-2.5-flare | gpt-image-2.5-sunburst | Swapping models is a one-string change in your code |
| Quality settings | low, medium, high, xhigh, max, auto | low, medium, high, xhigh, max, auto | Same ladder on both, two rungs higher than GPT Image 2 |
| Image tokens | $8.00 in, $30.00 out per million | $8.00 in, $30.00 out per million | Identical rates, so cost differences come from how many tokens each spends |
| Batch endpoint | Not supported | Not supported | High-volume jobs lose the batch discount GPT Image 2 still gets |
| Supported features | inpainting | inpainting | Both edit inside an existing image; neither lists anything else |
| Images per minute, Tier 1 | 5 | 5 | Same ceiling on how many images a minute you get, so Sunburst costs nothing extra in limits |
A dash (—) would mean OpenAI does not publish the figure; there are none in this table. Sources: https://developers.openai.com/api/docs/pricing and https://openai.com/index/introducing-chatgpt-images-2-5/ Checked 2026-09-09.
Independent developer Simon Willison noted on his blog on 8 September 2026:
He was quoting one line in OpenAI's own docs, which tells you to pick Sunburst when editing precision matters most. He then ran Sunburst on a reference image the same day, through a command-line tool he wrote himself.
Five quality settings, and what xhigh and max change
Previous OpenAI image models capped output quality at three distinct settings. Both 2.5 models support low, medium, high, xhigh, max and auto quality settings directly in the API. Both models default to auto if you don't declare a setting in your API call.
The standard sizes recommended for general production remain 1024x1024 pixels for square images, 1536x1024 for landscape, and 1024x1536 for portrait. Custom sizes work too, within four rules:
- Width and height must both be multiples of 16.
- The aspect ratio must sit between 1:3 and 3:1.
- Total pixel count must stay between 655,360 and 8,294,400.
- No single edge can pass 3840 pixels.
Any size above 2560x1440 is marked experimental.
ChatGPT's Thinking effort slider, shown at Medium. It isn't the API's quality setting — 2026-09-09.
You can also request transparent backgrounds by setting background: "transparent" alongside PNG or WebP output formats. Outputs default to PNG, but you can choose JPEG or WebP with compression quality scored between 0 and 100. JPEG formats return faster than PNG.
If you choose to stream generations in real time, OpenAI bills an extra 100 image output tokens for each partial frame delivered.
Moving up the ladder to xhigh and max doesn't change the per-token price on the rate card. Instead, higher quality settings make the model consume more output tokens to generate fine textures, subtle lighting passes, and cleaner edges. Your final billing per image rises because the total token count expands.
What GPT Image 2.5 costs to run
OpenAI bills GPT Image 2.5 at token rates of $8.00 per million image input tokens and $30.00 per million image output tokens. Cached image input tokens cost $2.00 per million, and text input tokens cost $5.00 per million.
These token rates match GPT Image 2 dollar for dollar.
But running 2.5 models can change your actual monthly expenses in two ways:
- Token consumption per image increases: The official GPT Image 2 calculator doesn't estimate GPT Image 2.5 token consumption. Choosing
xhighormaxoutputs spends more tokens than standardhighrequests. You can check your potential expenses on our GPT Image cost calculator. - The Batch API is disabled: Both
gpt-image-2.5-flareandgpt-image-2.5-sunburstlist thev1/batchendpoint as Not supported. Jobs using GPT Image 2 Generator can get a 50% discount on non-urgent tasks via the batch queue, but 2.5 right now needs standard real-time pricing for every request.
Rate limits are identical across both models and scale strictly by your organization's usage tier:
- Tier 1: 5 images per minute, which is one image every twelve seconds.
- Tier 2: 20 images per minute.
- Tier 3: 50 images per minute.
- Tier 4: 150 images per minute.
- Tier 5: 250 images per minute.
Some developer accounts may also need Organization Verification before the API allows requests to the GPT Image family.
Where the AI Overview is wrong about GPT Image 2.5
Automated search summaries published several errors on launch day. Here's what search overviews got wrong, alongside the documented facts we verified on 2026-09-09:
- How long it takes: The Overview claimed Flare is "50% faster than GPT-Image-2". OpenAI's launch announcement stated that Flare achieved "up to 50% lower latency" compared to Images 2.0. Cutting how long it takes by up to 50% represents an upper ceiling under best conditions, not a universal doubling of generation speed across the board.
- Missing pricing: The Overview stated no pricing information for the new release. OpenAI's official rate card confirms that token pricing matches GPT Image 2 at $8.00 per million input tokens and $30.00 per million output tokens.
- Quality settings: The Overview failed to mention quality tiers entirely. Both 2.5 models introduce
xhighandmaxsettings, expanding the previous three-tier structure into five separate options plus anautodefault. - Batch support: The Overview made no mention of batch processing. In documentation checked on launch day, both 2.5 models show the
v1/batchendpoint as Not supported, meaning the 50% batch discount isn't available on 2.5.
What GPT Image 2.5 still gets wrong
While GPT Image 2.5 brings noticeable visual upgrades, OpenAI outlines four primary technical limitations in its developer documentation:
- Wait times on dense prompts: Complex prompts may take up to 2 minutes to process.
- Text rendering issues: It can still get text in the wrong place, or make it hard to read.
- Character and brand consistency: It may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations.
- Structural layout placement: It can struggle to put things exactly where you asked in a structured layout.
Precision limits also apply to inpainting (you choose part of an image and describe what should replace it) in ChatGPT. OpenAI notes in its help documentation: "Highlights are not always precise, and edits may extend beyond the area you selected."
OpenAI also highlighted new safety considerations in its official system card. The company noted that 2.5 allows for heightened realism that could, absent safeguards, allow more convincing deepfakes, including political, sexual, or otherwise sensitive imagery of real people, places, or events.
How to start using GPT Image 2.5 today
You can access ChatGPT Images 2.5 immediately across web browsers, iOS and Android apps, and Codex on all ChatGPT account tiers. Advanced Images with thinking features remain limited to Plus, Pro, and Business tiers, with enterprise and education rollouts planned for later.
Developers can call gpt-image-2.5-flare and gpt-image-2.5-sunburst directly through the official image generation and edit endpoints. If you want to compare outputs across multiple models without managing multiple API keys, you can run both Flare and Sunburst today on our GPT Image 2.5 page.
GPT Image 2.5 FAQ
What can GPT Image 2.5 do?
GPT Image 2.5 generates new visual files from text prompts, modifies existing pictures through conversational multi-turn instructions, and handles inpainting edits within designated image areas. It supports transparent backgrounds, custom pixel dimensions up to 3840 pixels per edge, mobile sketching, and pre-built design templates.
How much does GPT Image 2.5 cost per image?
OpenAI doesn't bill GPT Image 2.5 at a flat fixed cost per image. Billing is calculated on per-token consumption at $8.00 per million image input tokens and $30.00 per million image output tokens. Higher quality tiers like xhigh and max consume more tokens during generation, which raises the final cost per image compared to standard generations.
Is GPT Image 2.5 better than GPT Image 2?
GPT Image 2.5 gives you shorter wait times, two extra quality tiers (xhigh and max), transparent background options, and improved how well it follows the prompt during multi-turn edits. But GPT Image 2 remains more cost-effective for large non-urgent workloads because it supports the Batch endpoint at a 50% discount.
Which OpenAI image model should you pick?
Choose Sunburst if how you work needs premium picture generation, fine brand detail, or precise multi-turn editing instructions. Choose Flare if you need fast turnarounds for social media imagery, rapid concept prototyping, or high-volume everyday graphics.
The verdict
GPT Image 2.5 is a real upgrade, and a narrow one. You get two more quality settings and faster generations across standard chat and API interactions.
If you build interactive tools or need refined conversational edits, moving your prompts to Sunburst on our GPT Image 2.5 page is a natural upgrade. But if your system relies on OpenAI's Batch API for halved token rates, keep your automated batches on GPT Image 2. Batch generation is also coming to Seadanse.
We read OpenAI's model pages, rate card and system card on 2026-09-09 to put together this review.
Related articles
- GPT Image 2.5 vs GPT Image 2 — whether the upgrade is worth it for what you already build.
- GPT Image 2.5 Flare vs Sunburst — which of the two API models to call, and why the docs will not tell you.
- Is GPT Image 2.5 Free? — what every ChatGPT tier actually gets.
