GPT Image 2.5, documented end to end
OpenAI shipped ChatGPT Images 2.5 and two API models on September 8, 2026. This is an independent reference for what changed, what it costs, and how to get good output from it.
Try an image model right here
A live image-to-image playground, embedded below. It is not GPT Image 2.5 itself — OpenAI serves that through ChatGPT and its own API — but it is a free way to practise the prompt structure this site teaches before you spend anything. Upload a picture, describe the change, and watch how specific wording changes the result.
Third-party demo hosted on Hugging Face Spaces. It may sleep when idle — give it a moment to wake up. Output and availability are the operator’s, not ours.
Sep 8, 2026
Release date
2 models
New API tiers
-50%
Generation latency
6 tiers
Quality settings
What GPT Image 2.5 is
GPT Image 2.5 is the image generation and editing model OpenAI released on September 8, 2026.
It arrived on two surfaces at once. Inside ChatGPT it is branded ChatGPT Images 2.5 and rolled
out the same day to every tier, plus ChatGPT Work and Codex, on desktop, mobile and web. In the
API it arrived as two separate models, gpt-image-2.5-flare and
gpt-image-2.5-sunburst, which is the first time this family has
split its API lineup by workload rather than by size.
The predecessor, GPT Image 2, shipped on April 21, 2026 with 2K output, multiple aspect ratios, web research for current information and a split between Instant and Thinking modes. Five months later this release does not chase a new headline capability. It goes after the three complaints that dominated feedback on the previous generation: generation was slow, subjects morphed between edits, and asking for one change quietly rewrote parts of the frame nobody had touched. Each of those has a named fix in this release, which is why the version number moved by half a point rather than a whole one.
Scale is the reason any of this matters commercially. OpenAI says its stack now produces more than three billion images a week across ChatGPT and the API, so a fifty percent latency cut is not a benchmark line — it is a serving-cost decision that also happens to change how people work. When a render lands in a few seconds you iterate inside the conversation. When it takes half a minute you write one long prompt, submit it, and go do something else.
Start where you need to
Features
Nine changes, split into model behaviour and ChatGPT interface work.
API models
Flare and Sunburst: IDs, quality tiers, endpoints and request shapes.
Pricing
Token rates, per-image arithmetic, and how to stop a bill running away.
Prompt patterns
Four habits that fix most bad output, plus eight copy-ready prompts.
How to use it
Ten steps in ChatGPT, from first prompt to a clean edit chain.
vs GPT Image 2
What five months of iteration actually changed, line by line.
vs Nano Banana Pro
Where Google still wins, and which asset types flip the answer.
Full FAQ
Twenty questions, from release date to watermarking to migration.
Two API models, one decision
OpenAI's framing is unusually blunt: start with Flare, move to Sunburst when precision across edits is the binding constraint. The two models share token rates and quality ladders, so the choice is about latency against control, not about budget.
Default choice for most applications
Flare
gpt-image-2.5-flare Fast, high-quality everyday image generation
- Speed
- Very fast
- Performance
- Higher
- ›Creator and social content
- ›In-product image generation
- ›Visual search and thumbnails
- ›Rapid image prototyping
- ›High-volume batch jobs
Delivers higher-quality images than GPT Image 2 at 50% lower latency. Manus reported 2-4x the speed of GPT Image 2 in its own evaluations.
Premium visual workflows with tight edit control
Sunburst
gpt-image-2.5-sunburst Most capable model for generation and editing
- Speed
- Medium
- Performance
- Highest
- ›Production-ready campaign creative
- ›Polished product imagery
- ›Multi-turn retouching chains
- ›Brand-controlled art direction
- ›Anything going to print
Built for workflows where editing precision matters more than turnaround. Generation takes longer than Flare by design.
Full endpoint list, quality settings and request examples live on the GPT Image 2.5 API page.
What changed, in nine parts
Six of these are model behaviour and reach the API. Three are ChatGPT interface features and do not. Confusing the two is the most common way to plan a workflow that cannot ship.
Reference-photo fidelity
Working from a reference photo, subjects stay recognisable across new settings, styles and compositions. Lighting and texture read more naturally, and distinctive features are more likely to survive the transfer. This is the change OpenAI leads with, and it is the one that matters for anyone putting a real person, pet or product into a generated scene.
Scoped editing
Ask for one change and you get one change. The model is better at leaving untouched regions alone, even when subject and background are visually complex — swap a jacket while the pose holds, replace a backdrop while the product geometry stays put, rewrite copy without the layout shifting underneath it.
Edits that survive the thread
In long conversations, earlier edits carry forward instead of quietly decaying. Each new instruction builds on the last rather than re-rolling the image, which is what makes a ten-turn retouching session viable instead of a race against drift.
Up to 50% lower latency
Generation latency drops by up to half against Images 2.0. Autoregressive image generation has been the slow part of this model family since the beginning, so halving it changes the interaction pattern: you iterate rather than wait, submit, and go do something else.
Complex layouts and transparency
Denser layouts hold together, and transparent-background generation improved enough that Manus called it out specifically. For anyone producing overlay assets, sticker sheets or UI art, a clean alpha channel out of the box removes an entire cutout step.
Grounded real-world content
Images that contain real-world information render more accurately, which shows up most in infographics, charts, maps and anything with a factual label on it. The system card also flags improved infographic accuracy and layout as a headline capability.
Sketch
Type @Sketch in ChatGPT and draw directly in the app. The drawing becomes a reference for generation, which is the fastest route for anything easier to draw than describe — a layout, a pose, a rough spatial arrangement of objects.
Templates
Starting scaffolds for common formats such as posters and merch. You pick a template, then add your own message, design elements and style. It is a cold-start fix rather than a model capability, and it removes the blank-prompt problem for people who do not write prompts for a living.
Shareable prompts
A generated image can carry the exact prompt that produced it, so someone else can run the same idea against their own photos. It turns a finished image into a reusable recipe instead of a dead end.
Cost per image, estimated
Billing is token-based, so there is no official price per picture. GPT Image 2.5 carries the same token rates as GPT Image 2, which makes the previous generation's measured per-image figures the best available working estimate.
| Quality | 1024×1024 | Portrait / landscape | Typical use |
|---|---|---|---|
| low | ~$0.006 | ~$0.006 | Drafts, thumbnails, A/B sweeps |
| medium | ~$0.053 | ~$0.041 | Client-facing everyday work |
| high | ~$0.211 | ~$0.165 | Marketing and hero assets |
| xhigh | not published | not published | New in this generation |
| max | not published | not published | New in this generation |
Who shipped with it on day one
“The GPT-Image-2.5 models are available in Adobe Firefly, Adobe's creative AI studio.”
“Flare delivered high-quality images at two to four times the speed of GPT Image 2, with improved transparent-background generation.”
“The model preserves an original image's character, composition and visual identity through edits.”
Day-one distribution through Adobe Firefly is the more telling detail. A model that lands inside an established creative suite on release day has been in partner hands for weeks, which is consistent with the anonymous checkpoints spotted in public arena testing through August 2026.
Eight ways people waste money on this model
✗ Generating drafts at high quality
✓ Fix: Iterate at low, then re-run the winning prompt at high. The tier spread is roughly 35x per image, and you are discarding the draft anyway.
✗ Stacking four edits into one turn
✓ Fix: Scoped editing is per-instruction. Split them: one change, then the next, and the thread will carry both.
✗ Leaving text unquoted
✓ Fix: Quote every string you want rendered. Unquoted words get paraphrased into something that looks like text but is not.
✗ Cropping to a ratio after the fact
✓ Fix: Ask for the ratio in the prompt or the picker. A post-hoc crop throws away composition the model spent its budget on.
✗ Reaching for Sunburst by default
✓ Fix: Flare is the default for a reason. Sunburst buys edit precision at the cost of generation time; pay that only when the asset ships.
✗ Assuming a per-image price
✓ Fix: Billing is token-based. Size and quality both move the number, so estimate before you queue ten thousand jobs.
✗ Describing a mood instead of a frame
✓ Fix: 'Cinematic and moody' is not an instruction. Name the light source, its direction, and what it falls on.
✗ Restarting the thread after a bad edit
✓ Fix: Ask for the previous state back instead. The conversation is the edit history, and losing it costs you every earlier refinement.
Safety and provenance
OpenAI published a system card for this release alongside the announcement. It describes the safety stack as built on the same foundations used for Images 2.0, with additional safeguards for risks that emerge as the model gets more capable. Under the Preparedness Framework, both Sunburst and Flare were evaluated for biological and cybersecurity capability by asking the model to render a reasoning scratchpad and a final answer inside an image, then grading the answer. Neither crossed the Bio High or Cyber High thresholds.
AI Self-Improvement is not tracked for this family, on the reasoning that an image model cannot write and execute code in a way that would enable it. That is a defensible line for now, though it sits slightly awkwardly beside a release whose headline is more accurate rendering of real-world information, including text and diagrams.
On provenance, images generated in ChatGPT in this family have carried C2PA content credentials. If your workflow depends on that metadata surviving, verify it on your own output rather than assuming it, because handling differs between the chat surface and API responses.
GPT Image 2.5 FAQ
What is GPT Image 2.5?
Is GPT Image 2.5 actually available now?
What are the two API model IDs?
How much faster is it?
Does GPT Image 2.5 cost more than the previous model?
What is Sketch?
Can I use the output commercially?
Is this site run by OpenAI?
This page tracks a moving target
GPT Image 2.5 is days old. Tier pricing, resolution ceilings and endpoint behaviour will firm up over the coming weeks, and this reference gets updated as they do. Bookmark it, or jump straight to the section you need.