How to Use GPT Image 2.5: Prompts, Models, Settings

The first thing I asked GPT Image 2.5 to do was change a jacket. Just the jacket — navy instead of red, everything else untouched. Older models had turned that request into a lottery: the jacket changed, and so did the face, and by the third attempt I had lost the version I actually liked. This time the jacket changed and nothing else moved. That single test is most of what is new here, and most of what this guide is about.

GPT Image 2.5 shipped on 8 September 2026 US time (the small hours of 9 September in Beijing) under the product name ChatGPT Images 2.5. If you want the release-date facts and the model lineage, that is it. The rest of this page is how to actually drive it.

What you are choosing between

The API split into two engines for the first time. Everything else about them is the same — same parameters, same inputs, same outputs — so the only real decision is which one you point at.

Flare is OpenAI's default, and it should be yours too. One published test clocked it at two to four times the speed of GPT Image 2. You will spend most of your session in Flare and switch to Sunburst for the last two or three passes, when you already know what the picture should look like and you are protecting fine detail rather than exploring.

If you have seen people searching for GPT Image Pro, Sunburst is the thing they are describing. There is no separate "Pro" model.

The previous generation, GPT Image 2, is still available and has not been retired. It is a reasonable fallback if you have prompts tuned to it.

Quality tiers, and the one that trips people up

Five tiers: low, medium, high, xhigh, max. GPT Image 2 capped at high, so xhigh and max are genuinely new.

The tier people get wrong is auto — because there is no auto quality. auto is a value of the size parameter, not the quality parameter. Send quality: "auto" and the request fails. That distinction is worth committing to memory because the two parameters sit next to each other in every example you will read.

Alongside quality there is a separate resolution tier — 1K, 2K, 4K — which is ignored when you have set size to auto or to explicit pixel dimensions.

What to actually pick. Start at medium + 1K. It is fast and cheap enough that you can iterate without thinking about it. Move to high when the composition is settled. Reach for xhigh or max only for a final asset that is going to print or to a client, because the cost climbs steeply and the visible difference on a screen is small.

Two more settings worth knowing:

• background: "transparent" gives you a real alpha channel. The transparent output has visibly less white fringing than the previous generation did — that fringe used to show up the moment you dropped a cut-out onto a dark background.
• n accepts 1 to 10, so you can ask for a set in one call.

Three ways in, and what each one costs you

Before the settings, decide where you are running this.

The free ChatGPT tier is real but tight — testers report roughly two to three images a day, and hitting the ceiling can silently downgrade you to a lighter model. If your results suddenly get worse mid-session, that is usually what happened, not a bad prompt.

The five interaction surfaces

All of the new interaction lives in the ChatGPT conversation box. Nothing to install, no model dropdown to manage — 2.5 replaced 2.0 as the default.

Templates are the lowest-effort surface: pick one, upload a product photo, choose a style, choose a scene, and you never write a prompt at all. Comments do more than recolour — you can delete a background, erase an object or resize something inside the frame from the same interface.

These features are rolling out gradually, so an entry point missing from your account is usually a rollout gap rather than a fault.

Sizes and aspect ratios

size accepts auto, a named ratio, or explicit pixels.

The named ratios: 1:1, 3:2, 2:3, 4:3, 3:4, 4:5, 5:4, 16:9, 9:16, 21:9, 9:21, 2:1, 1:2, 3:1, 1:3.

For explicit pixels, dimensions must be multiples of 16 and total between roughly 655,000 and 8,300,000 pixels. Sending a ratio outside the supported set is the second most common cause of a failed request, after the quality: "auto" mistake.

Practical picks: 4:5 for anything going to a social feed or a product listing, 16:9 for a header, 1:1 when you want the model to spend its resolution on the subject rather than on empty margins.

Working with reference images

You can pass up to 16 reference images in one request. With more than one, the failure mode is the model averaging them into mush. The fix is to assign each a job explicitly:

Use image 1 for the scene and lighting.
Use image 2 for the subject's face and hair — preserve them exactly.
Use image 3 for the jacket only; ignore everything else in that image.

The strongest version of this pattern is the preservation list taken to its extreme, which is what makes virtual try-on work:

Dress the person using the provided clothing images. Do not change her face,
facial features, skin tone, body shape, pose, or identity in any way.
Preserve her exact likeness, expression, hairstyle, and proportions.
Replace only the clothing, fitting the garments naturally to her existing
pose and body geometry with realistic fabric behaviour. Match lighting,
shadows, and colour temperature to the original photo so the outfit
integrates photorealistically, without looking pasted on. Do not change the
background, camera angle, framing, or image quality, and do not add
accessories, text, logos, or watermarks.

Read the shape of that: a complete list of what must not move, then one sentence of what may, then an exclusion list. Every reliable edit prompt is that shape.

For photos of people, a shorter lock line covers most cases:

Keep my exact face, likeness and identity unchanged: do not alter features,
age, or skin tone.

Writing a prompt that survives

OpenAI published a prompting guide alongside the model. Stripped of the framing, it comes down to a handful of habits, and the two that matter most for this generation are the last two.

Name the deliverable first. Open by saying what kind of thing this is — a product photo, a poster, a diagram — then give composition, aspect ratio and constraints. A model that knows it is making a product photo makes better guesses about lighting than one that has only been handed a list of nouns.

Describe what is visible. Material, light, colour. "Photorealistic" works as an instruction. Camera settings are mood, not physics — writing f/1.4 nudges the look, it does not guarantee a depth of field.

Be specific about bodies. "Hands naturally gripping the handlebars" gets you something. "A person on a bike" gets you a lottery ticket.

Put required text in quotes. Anything that must appear verbatim goes inside quotation marks. This is the single highest-value habit for posters and product shots.

Separate what changes from what stays. When editing, write change only X and then list what must be preserved. This is the habit the whole generation is built around, and the next section is entirely about it.

Assign roles to reference images. With multiple inputs, number them and say what each one is for: image 1 supplies the scene, image 2 supplies the subject.

Change one thing per turn. Each round should carry one instruction, and should repeat the constraints you care about. "Same style as before" carries context, but when a constraint starts drifting you have to restate it in full.

A prompt skeleton you can paste

[Deliverable] of [subject], [aspect ratio].
Composition: [what goes where, what is in focus]
Look: [material, lighting, colour, mood]
Text (exact, verbatim): "[line one]" / "[line two]"
Do not: [things that must not appear — extra text, logos, watermarks]

This is deliberately dull. The interesting part of a prompt is the specificity you put into the brackets, not the structure around them.

Editing one region without wrecking the rest

This is the reason to use GPT Image 2.5 over what came before, so it is worth being precise about how to ask.

Comment on the region you want changed — mark it, then say what it should become. Everything outside the marked area is preserved rather than regenerated around your request. The instruction pattern that works:

In this photo, replace ONLY the white chairs with wooden chairs.
Preserve camera angle, room lighting, floor shadows, and surrounding objects.
Keep all other aspects of the image unchanged.

Three things are doing work there. The word ONLY, capitalised. An explicit preservation list. And a catch-all line at the end for everything the list forgot.

The shortest version of this that still works is genuinely short:

Remove the flower from the man's hand. Do not change anything else.

Region editing also covers more than colour changes: you can delete a background, erase an object, or change the size of something inside the frame.

The honest limits

Multi-turn editing is much better than it was. It is not solved. Independent hands-on testing found that a long chain of edits still produces light artefacts on faces, and that region isolation is not perfect — some pixels outside the marked area do shift. Treat it as a large reduction in re-rolling rather than the end of re-rolling.

The same testing found character consistency across separate images improved less than the other areas. If you are generating a character across four panels, expect to restate the character description in full every time, and expect to discard some outputs.

Working from a sketch

You can draw a rough composition and have the model finish it. The important finding here is quantitative, and it changes how you should use the feature.

A published test ran eight scenes through three conditions and scored the results out of 100:

Sketch alone was not just worse — it was unstable, with one case scoring 19. The lesson is blunt: a sketch supplies composition, the text supplies the destination. Handing over a drawing without a written brief is the worst of the three options you have.

A workable minimum:

Turn this rough sketch into a polished finished image.
Preserve the composition and every object. Add no text.

Then add the brief on top — name the subject, state the medium and palette, list what must not appear.

Text and non-Latin layouts

There is a persistent misconception that this generation is where Chinese characters started rendering properly. That is not accurate. GPT Image 2 already handled Chinese layouts well. What 2.5 adds is stability while editing them: change one headline in a dense poster and the several hundred surrounding characters stay where they were. Previously that edit was a coin flip for the whole layout.

Small type can still contain errors. Proof a dense poster before it goes to print — read the fine print, not just the headline.

The way to get text right is the habit from earlier: put every required string in quotes, list them one per line, and give the disclaimer text its own line rather than burying it in a paragraph of instructions.

When it fails, is slow, or runs out

Generation failed. The most common cause is a parameter the model does not accept. quality: "auto" is the classic. Sending an aspect ratio outside the supported set is the other one — check your ratio against the accepted list before assuming the model is at fault.

It is slow. You are probably on Sunburst. Switch to Flare while you explore; the difference is the whole point of the split. Very large sizes and the xhigh / max tiers also add real time.

You hit a limit. Inside ChatGPT, 2.5 is rolling out to all tiers including the free one, but OpenAI published no fixed quota. Testers report roughly two to three images a day on a free account, and — this is the part that catches people — hitting the ceiling can silently drop you to a lighter model without an obvious warning. If your output quality falls off a cliff mid-session, check whether you are still on the model you think you are.

Through the API there is no daily cap; you are billed per token instead.

What it costs

Token rates match the previous generation: image output at $30 per million tokens, image input at $8, text input at $5.

The catch is that identical rates do not mean identical cost per image. OpenAI did not publish how many tokens 2.5 burns per image, and quality tier and resolution move that number a lot. Budgeting from your GPT Image 2 spreadsheet will be wrong — probably not by a small margin. Measure your own first hundred images before you commit to a price.

A three-step routine

1. Draft on Flare at medium + 1K. Get the composition right. Do not touch quality settings yet; you are still deciding what the picture is.
2. Refine by pointing. One change per turn, with an explicit preservation list. Restate constraints that start drifting instead of relying on "same as before".
3. Final pass on Sunburst. Raise quality only now, and only for the version you are actually shipping.

Most of the wasted spend I have seen comes from doing step three first — generating at max while still exploring, then throwing the result away.

Try it

You can run GPT Image 2.5 on the generator on this site without a ChatGPT subscription or an API key, and the first render is free. If you would rather see what the model is good for before learning the controls, the companion piece walks through four real workflows — retouching, backgrounds, product shots and posters.

image25.online is an independent third-party platform and is not affiliated with OpenAI. Site credits are separate from any official account.