OpenAI shipped GPT Image 2.5 on September 8, 2026, and it is an unusually easy model to start with and an unusually easy model to use badly. Most wasted credits come from three habits: leaving quality on the top tier by reflex, asking for sizes the model quietly rounds off, and rewriting a whole prompt when one sentence needed changing. This page is the control manual — every setting, every size rule, the prompting pattern OpenAI actually documents, and the editing technique that keeps your fourth revision from undoing your first.
Sixty-second quick start
Do this once and the rest of the page will make sense.
- Open the ChatGPT Image 2.5. No ChatGPT subscription is required — it runs in the browser.
- Set the mode to Draft. Not Standard, not Ultra. You are testing your prompt at this stage, not the model.
- Set the size to 1K and pick your aspect ratio from the presets rather than typing custom pixels.
- Write a prompt with four parts in this order: subject, composition, style, lighting. One sentence each is enough for a first pass.
- Generate, then re-run the same prompt two or three more times. Draft runs cost 4 credits each, so four exploratory renders still cost less than a single Ultra image at 22.
- Read the results for composition problems, not detail problems — detail is what higher tiers fix, composition is what your prompt fixes.
- Once one render is structurally right, re-run that exact prompt at Standard or Ultra.
That last step is the whole method. Draft finds the prompt; the higher tiers render it. Reverse the order and you pay 22 credits to learn that your framing was wrong.
The settings that matter
GPT Image 2.5 ships as two API models — gpt-image-2.5-flare, the fast default, and gpt-image-2.5-sunburst, tuned for precision on detailed creative work. Per OpenAI's model reference, both share the same rate card, the same quality tiers, and the same two endpoints (generation and edits); the difference is speed versus precision, not capability. Both models succeed ChatGPT Images 2.0 from April 2026 and sit on the same autoregressive multimodal line.
Quality tiers
The API exposes six settings: auto, low, medium, high, xhigh, and max. The last two are new in 2.5 — 2.0 stopped at high. Cost scales with resolution multiplied by quality, so the same tier is not the same price at 1K and 4K. At 1024×1024, OpenAI's token pricing works out to roughly $0.006 per image at low, about $0.053 at high, and about $0.21 at max.
On this site those tiers are collapsed into three modes with fixed credit prices:
| Size |
Draft |
Standard |
Ultra |
| 1K |
4 credits |
6 credits |
22 credits |
| 2K |
5 credits |
8 credits |
24 credits |
| 4K |
6 credits |
14 credits |
42 credits |
Image-to-image costs the same as text-to-image at every cell in that table, which matters more than it sounds — it means iterating on an existing image carries no surcharge, so there is no financial reason to try to nail everything in one generation.
When Draft is genuinely enough. Draft is not a degraded preview. It is the correct final setting for a large share of real work:
- Thumbnails, social avatars, and anything that will be displayed under about 600px.
- Backgrounds and textures that sit behind text or UI.
- Concept exploration where you are choosing between directions, not shipping.
- Any image you plan to edit two or three more times — spend the credits on the final pass, not the intermediate ones.
- Blog headers and card images at 1K, where compression at delivery erases most of what Ultra bought you.
When to pay for Ultra. Dense in-image text, multi-font layouts, diagrams, product shots where material texture is the point, anything printed, and faces at large display sizes. OpenAI's own guidance is to raise quality specifically for dense text and diagrams — the tier buys glyph accuracy and fine texture, not better composition or better prompt following.
The rule that saves the most credits: do not assume a higher tier improves every prompt. Test the lowest setting that clears your acceptance bar, then stop. A simple flat-illustration logo concept looks identical at Draft and Ultra; you would be paying 18 extra credits for a difference that does not exist in that style.
background accepts auto, opaque, or transparent. Transparency only survives in PNG or WebP — request a transparent background with JPEG output and you get a solid fill, usually white, with no warning. If you are making stickers, logos, overlays, or UI assets, set the format before you set anything else.
Worth being clear about: transparent backgrounds are not new in 2.5. They are family-wide across the GPT Image line, as are the reference-image and size limits below. What is genuinely new in 2.5 is the xhigh and max tiers, the Flare/Sunburst split, the latency improvement OpenAI claims over GPT Image 2, and edit stability across turns.
Sizes and aspect ratios
These are hard validation rules, not suggestions. Requests that violate them fail or get silently adjusted.
- Both edges must be multiples of 16. 1024, 1536, 2048, 3840 all work. 1000 and 1500 do not.
- Maximum edge: 3,840 pixels. No side may exceed it, in either orientation.
- Aspect ratio must sit between 1:3 and 3:1. Panoramas wider than 3:1 and skyscraper banners taller than 1:3 are out of range. Generate at 3:1 and crop, or extend in an editor.
- Total pixels: 655,360 to 8,294,400. The floor rules out very small custom sizes; the ceiling is what 3840×2160 lands just under.
- Anything above 2560×1440 is experimental. OpenAI flags it as such. Expect more variance in structure and occasional artifacts at the very top of the range — this is the one setting where paying more can produce a worse result.
Safe sizes to use without thinking: 1024×1024 (1:1), 1536×1024 (3:2 landscape), 1024×1536 (2:3 portrait), 2048×2048, and 3840×2160 (16:9, experimental territory).
Two practical notes. First, aspect ratio belongs in the setting, not only in the prompt — writing "16:9 widescreen" in prose while the size field says 1024×1024 produces a square image with letterbox bars painted into it. Set the field, then optionally reinforce it in the prompt. Second, if you need an odd ratio like 4:5 for a feed post, compute it against the multiple-of-16 rule first: 1024×1280 works, 1024×1279 does not.
Prompting that works
OpenAI's image-prompting guide describes a specific hierarchy, and following it beats creative phrasing almost every time. The order is: subject and intent → composition → style, materials, and lighting → people and action → exact quoted text → constraints. For complex requests the guide recommends organizing the prompt into labeled sections rather than one long paragraph.
Three rules carry most of the weight:
- Name the artifact, not just the content. "Product photograph," "technical diagram," "editorial illustration," "advertisement" — the category word sets conventions the model already knows.
- Put required wording in quotation marks, with position and typography specified. The model renders quoted strings as literal text; unquoted text becomes a suggestion it may paraphrase or misspell.
- State constraints as constraints. A closing line listing what must not appear or change does real work. This is also the mechanism that makes editing reliable.
Four prompts you can copy and adapt:
Product shot
Product photograph of a matte-black stainless steel water bottle, centered, three-quarter angle, occupying about 60% of the frame height. Studio setup on a seamless warm-grey backdrop. Soft large key light from upper left, subtle rim light on the right edge, gentle contact shadow beneath. Photorealistic, shallow depth of field. Constraints: no reflections of studio equipment, no visible branding, no text anywhere in the image.
Text-forward poster
Editorial poster, portrait orientation. A single ripe pomegranate split open, bottom third of the frame, dramatic side lighting from the right. Flat matte background in deep ink blue. Headline text reading "SEASON OF SEEDS" in a bold condensed sans-serif, centered in the upper third, cream white. Smaller line reading "November 14 to 30" directly beneath in the same family, letter-spaced. Constraints: exactly these two lines of text and no other lettering, no drop shadows on the type.
Diagram
Technical diagram, clean vector style, 3:2 landscape. Three labeled boxes arranged left to right connected by arrows: box one reading "Draft", box two reading "Review", box three reading "Publish". Thin 2px strokes, single accent color on the arrows, everything else greyscale. White background. Constraints: no gradients, no icons inside the boxes, no additional labels beyond the three named.
Character illustration
Editorial illustration of a woman in her sixties repairing a bicycle wheel on a city sidewalk, half-body framing, seen from slightly below, her gaze down at her hands. Loose gouache texture with visible brush edges, limited palette of ochre, teal, and off-white. Late afternoon light raking from the left. Constraints: keep hands clearly visible and correctly formed, no text, no background crowd.
Notice what those share: every one ends with constraints, every quoted string is the exact text wanted, and none of them use vague quality adjectives like "beautiful," "masterpiece," or "8K." Those words consume prompt space without changing output — resolution and quality are settings, not adjectives.
The prompt library has more structured examples if you want starting points rather than blank-page phrasing.
Editing without wrecking your image
This is where 2.5 differs most from 2.0, and where most people still lose good images. OpenAI's editing pattern is a single sentence: "change only X," followed by a list of what to preserve.
The technique, turn by turn:
Restate preservation every single turn. Not just the first. Constraints drift as the conversation grows, and the fix is repetition, not cleverness. If the lighting, camera angle, color grade, and background matter, say so again on turn four exactly as you said it on turn one. OpenAI's guide is explicit that critical constraints should be restated across edit turns to prevent drift.
One change per turn. Two changes in one instruction means you cannot tell which one caused a regression, and it roughly doubles the chance the model re-renders regions you wanted untouched. Three sequential single-change edits at Draft cost 12 credits and are more controllable than one compound edit at Ultra.
Pass the previous output forward. Each edit takes the last result as input, not the original. That is what accumulates changes instead of restarting from scratch.
A working edit instruction looks like this:
Change only the jacket color, from navy to burnt orange. Keep the pose, facial features, hair, lighting direction, background, camera angle, color grade, and the fabric's visible texture exactly as they are. Do not alter the composition or crop.
Reference images and roles. Up to 16 reference images are supported, each under 50MB. The rule that makes multi-reference work: label every input by number and purpose, then explain how they combine. Unlabeled references get averaged into mush.
Image 1 is the subject — use this person's face and build. Image 2 is the style reference — copy its color palette, grain, and lighting quality, not its subject or composition. Image 3 is the wardrobe — reproduce this jacket's cut and material. Place the subject from image 1 wearing the jacket from image 3, rendered in the style of image 2, in a plain studio setting.
Most reference-image failures are role failures. If you upload three photos and write "combine these," the model has to guess which one supplies identity, which supplies style, and which supplies layout — and it will guess differently every run.
Know when to restart. After roughly five or six edit turns, drift accumulates faster than you can correct it. When two consecutive edits each fix one thing and break another, stop editing. Take the best version you have, describe it as a fresh prompt with your constraints baked in, and generate again.
If you're in ChatGPT instead
The 2.5-era ChatGPT release added several product features that have no API equivalent:
- @Sketch — draw a rough layout and have the model build from your marks.
- Templates — reusable saved setups for repeated formats.
- Comment-based edits — leave a note on a specific region and have that region changed.
- Prompt sharing — pass a working prompt to someone else directly.
These are ChatGPT product features, not model capabilities, and they are not exposed through the API. Images 2.5 reached all ChatGPT tiers including free at launch, subject to rate limits. There is no Batch API support for 2.5 at launch, so high-volume pipelines lose the discount they had on the older model.
If you want the model's actual output without a subscription, the browser tool on this site runs the same generation and editing path with credits instead of a monthly plan. The free weekly check-in gives 30 credits, which covers about seven Draft 1K images or five Standard ones — enough to test whether the model fits your work before spending anything.
Five mistakes that waste credits
1. Starting at Ultra. The most expensive habit there is. Ultra at 1K costs 22 credits; the same prompt at Draft costs 4. Prompt problems are visible at Draft. Run 4 Draft images, fix the prompt, then spend once.
2. Rewriting the whole prompt to fix one thing. If the composition is right and the color is wrong, edit the color. Regenerating from a rewritten prompt throws away everything that worked and re-rolls the parts you liked.
3. Requesting sizes that violate the rules. Typing 1000×1000 or a 4:1 banner produces silent adjustment or a failed request. Check the multiple-of-16 and 3:1 rules before you type a custom size.
4. Generating at 4K "to be safe." 4K Ultra costs 42 credits, sits in experimental territory above 2560×1440, and is wasted entirely if the image ends up in a 1200px-wide blog post. Generate at the size you will actually publish.
5. Skipping the preservation list on edits. An edit instruction with no constraints invites the model to re-render regions you were happy with. The list costs you one sentence and saves the re-do.
A sixth, less common but expensive: requesting a transparent background with JPEG output. Transparency requires PNG or WebP. You will pay full price for an image with a white rectangle where the transparency should be.
FAQ
Should I use Flare or Sunburst?
Flare is the fast default and the right choice for iteration, high-volume work, and anything where you will run the prompt several times. Sunburst trades generation time for precision on detailed creative work and exacting edits. They share pricing and quality tiers, so the decision is purely speed versus precision — start on Flare and move up only when you can point at a specific detail Flare missed.
What is the difference between Draft, Standard, and Ultra here?
They map onto the underlying quality tiers at fixed credit prices: 4, 6, and 22 credits at 1K, rising to 6, 14, and 42 at 4K. Draft is for finding the prompt and for any output displayed small. Standard covers most finished work. Ultra is for dense text, fine texture, faces at large sizes, and print.
What image sizes does GPT Image 2.5 support?
Both edges must be multiples of 16, no edge may exceed 3,840 pixels, the aspect ratio must fall between 1:3 and 3:1, and total pixels must land between 655,360 and 8,294,400. Anything above 2560×1440 is flagged experimental. Reliable choices are 1024×1024, 1536×1024, 1024×1536, and 2048×2048.
How many reference images can I use at once?
Up to 16, each under 50MB. The practical limit is lower — beyond four or five, assigning a clear role to each becomes the bottleneck. Label every reference by number and purpose ("image 1 = subject, image 2 = style") and state how they should combine, or the model will average them.
How do I get text inside the image to render correctly?
Put the exact wording in quotation marks and specify position and typography around it. Unquoted text is treated as a description the model may paraphrase. For dense text, multi-font layouts, or anything with more than a couple of lines, OpenAI's guidance is to raise the quality setting — this is one of the few cases where the top tiers earn their cost outright.
Why did my edit change things I did not ask about?
Because the instruction did not say what to preserve. Use the "change only X, keep Y" pattern and list the specifics — pose, lighting, background, camera angle, color grade — and restate that list on every turn, not just the first. Also make one change per turn; compound edits are much more likely to re-render regions you wanted untouched.
Do I need a ChatGPT subscription to use GPT Image 2.5?
No. Images 2.5 reached all ChatGPT tiers including free at launch with rate limits, and the browser tool on this site runs the model on credits with no subscription at all. The free weekly check-in provides 30 credits; packs start at $9.90 for 80 credits and never expire, and the $29.90 plan includes 300 credits with a commercial license and no watermark — details on the pricing page.
Open the GPT Image 2.5 image generator, set the mode to Draft, and run one prompt four ways before you spend anything on Ultra. The habit is worth more than any prompt in this guide.
Sources
- Image prompting — OpenAI API docs — verified 200: the official prompt hierarchy (subject, composition, details, quoted text, constraints), the "change only X" edit pattern, reference-image roles, multi-turn iteration, and the note that custom sizes must be multiples of 16 with a 3,840px maximum edge.
- GPT-Image-2.5 Flare model reference — OpenAI — verified 200: the
gpt-image-2.5-flare-2026-09-08 snapshot, the six quality tiers including xhigh and max, token pricing, and the generation/edits endpoints with batch unsupported.
- GPT-Image-2.5 Sunburst model reference — OpenAI — verified 200: matching rate card and quality tiers, confirming that the Flare/Sunburst difference is speed versus precision rather than pricing or capability.
- Create image — OpenAI API reference — verified 200: the
quality, size, background, and output_format parameters, and the transparency requirement for PNG or WebP output.
- Introducing ChatGPT Images 2.5 — OpenAI — the September 8, 2026 launch announcement covering Flare and Sunburst positioning, latency claims, subject preservation, and the ChatGPT-side features. Page blocks non-browser requests; confirmed through search indexing rather than direct fetch.
