How to prompt GPT Image 2.5—and why the quality tier costs more than the size
On GPT Image 2.5, prose and reference images are the whole control surface: no seed, no negative-prompt field, no style dial. The quality tier moves the price far more than the pixels do—a 4K image at medium quality quoted less than a 1024-pixel one at high.
GPT Image 2.5 comes in two models, and OpenAI is direct about which is which: "Choose Sunburst for workflows where editing precision matters most, and Flare for fast, high-quality everyday image generation." On LUVI both run as text-to-image and edit, next to the older GPT Image 2 and 1.5. On September 22, 2026 a 1024×1024 image at the medium tier quoted 30 credits. One thing to know before you start: GPT Image 2.5 has its own prompting manual in LUVI, and OpenAI publishes a separate 2.5 prompting guide. Neither carries over the GPT Image 2 rules on three points: any prompt format works, plain negatives are safe, and edits go one change at a time.
What order does the prompt go in?
The deliverable and its use come first, in the opening words: "a vertical retail poster for a coffee roastery", "a product photograph for a web shop". OpenAI's 2.5 guide opens the same way: "Start with the image you need, then describe the subject, composition, style, and constraints." For a long brief, LUVI's 2.5 manual uses the order scene, subject, details, constraints, and labeled sections such as "Scene:", "Subject:", "Style:", "Text:" and "Constraints:" are fine.
Then describe what is visible: materials, light, colour, medium, framing. "Warm tungsten key from the left, shallow depth of field" beats "cinematic lighting", and a lens is a look cue, not physics. OpenAI's words: "Treat camera specifications as cues for appearance, not a guarantee of exact physical simulation." Don't add what the brief lacks, either: the model already fills empty scenes with extras of its own.
The format is free. "Short prompts, descriptive paragraphs, JSON-like structures, instructions, and tags can all express the same intent," says OpenAI's 2.5 guide, so the old GPT Image 2 rule of prose over tags doesn't apply here.
Which words can I leave out?
Quality boosters such as "8k", "masterpiece", "best quality" or "ultra-detailed". LUVI's 2.5 manual says to replace them, and mood words like "cinematic" or "stunning", with scale, light, colour and material, because boosters steer less than concrete detail. In LUVI's test renders for that manual they did no harm and no good. Leave the shape out of the sentence too: "16:9" or "4K" in the prompt sets nothing, because the size control does that job.
You may have read about a speckled "tiling" artifact on foliage and fur, with a list of anti-speckle words to fight it. That guard belongs to GPT Image 2. In LUVI's test renders for the 2.5 manual, dense grass, olive foliage and clouds came back without it, so the extra vocabulary isn't needed.
Do negative instructions work?
Yes. On GPT Image 2.5, plain negatives are safe and obeyed, the opposite of the GPT Image 2 advice. LUVI's manual says to end the prompt with the exclusions the result needs, written as short plain negatives such as "No other text." and "No watermark, no logos." OpenAI's own 2.5 example prompts close the same way: "No extra text, no watermarks, no unrelated logos."
One exclusion is worth adding by default. When the scene invites lettering you didn't ask for (a named town, a harbor, a street, a shopfront, a vehicle), end with "No text, signs or lettering anywhere."
How do I get a transparent background?
With the parameter, never with the prompt. Asking for a transparent background in prose paints a literal checkerboard as solid pixels—the model draws the thing that represents transparency instead of removing the background. LUVI's own parameter contract carries that warning next to the control.
OpenAI's instruction is the short version: "For transparent backgrounds with either model, set background: "transparent" and use output_format: "png"". On LUVI those are two dropdowns, background and output_format, and setting them is the entire job.
What actually moves the price?
The quality tier, by a long way—and far more than the number of pixels.
- 1024×1024, medium: 30 credits
- 1024×1024, high: 93 credits
- 1024×1024, max: 346 credits
- 3840×2160, medium: 50 credits
Read those together and the shape is clear: going all the way to 4K at the medium tier costs less than staying at 1024 pixels and moving one tier up. xhigh and max are new to 2.5—OpenAI notes that earlier GPT Image models stop at high—and they are where the credits go.
A practical habit follows. Draft at low or medium, decide the composition, and only then spend a tier. OpenAI suggests the same: "Use quality: "low" for quick drafts."
Sizes come from a fixed list. On LUVI, size offers 14 sizes, from 1024×768 up to 3840×2160 in both orientations, and it is the only control over dimensions and aspect ratio; there is no free width and height. OpenAI marks resolutions above 2560×1440 as experimental.
Is the estimate what I actually pay?
No—it is the ceiling, and on this family the gap is visible. We generated seven images at 2048×1152, medium quality, JPEG output. One of them quoted 26 credits before the run; across all seven, the amounts actually taken off the account ranged from 19 to 21.
That is how charging works across LUVI: what you see before you run is the most you will pay, and the charge settles at the provider's real bill. It just happens to be easy to see here, because this family's estimate runs consistently above its settled price.
What changes in the edit models?
The prompt becomes an instruction about a picture that already exists, and LUVI's 2.5 manual sets the order: the change first, stated exactly, then what must hold. "Change only the jacket to a red denim jacket. Keep everything else exactly the same: her face, her pose and the room lighting." Restate the full keep list every time rather than assuming silence protects anything.
Make one change per edit. The manual puts it plainly: "OpenAI's advice for chains is one change per request, restating what to keep." OpenAI's 2.5 guide says "request one change, and repeat the details to preserve", so pass the last output back in as Image 1 and make the next change from there. This reverses the GPT Image 2 advice to batch every change into one prompt.
With several inputs, refer to them as Image 1, Image 2 and so on in upload order, and give each a role and a noun: "Image 1: base scene to preserve. Image 2: jacket reference. Dress the person from Image 1 in the jacket from Image 2." The edit models on LUVI take up to 16 images and have no mask. Every edit re-renders the whole frame, so the keep clause does the real work; if a region must stay pixel-identical, composite it back into the original afterwards.
How we tested
- Models: GPT Image 2.5 Flare Text-to-Image, with Sunburst, GPT Image 2 and 1.5 read from the catalog.
- Settings: 1024×1024 at medium, high and max; 3840×2160 at medium; and seven real generations at 2048×1152, medium, JPEG.
- Date and runs: September 22, 2026—seven generations, plus live credit quotes and the models' parameter contracts in LUVI, alongside OpenAI's image generation guide.
- Results: quotes of 30, 93 and 346 credits across the three tiers at 1024×1024, and 50 credits at 3840×2160 medium. The seven real images were charged between 19 and 21 credits each, against a 26-credit quote for one of them.
- Re-checked: October 20, 2026. The same credit quotes came back, and we read the four models' parameter contracts again along with LUVI's GPT Image 2.5 manual, which the prompting advice above follows.
- Drawbacks: the seven images were all at one setting and one kind of subject, so the estimate-to-charge gap is measured for that cell only, not across the whole size and quality grid.
In LUVI
Open Flare or Sunburst in the Workspace, open the prompt with the deliverable and close it with the exclusions, and set background and output_format in the controls rather than in the sentence. Check the credit estimate above the Generate button before you raise the quality tier—it is where the cost of this family lives, and the help center explains how resolution and duration change the price more generally.
New accounts start with 200 credits, and verifying a phone and a card adds up to 1,500 more. You can create a LUVI account and find the tier your work actually needs.
Sources
- Image generation guide — OpenAI, official platform documentation. Accessed October 20, 2026.
- GPT Image 2.5 prompting guide — OpenAI, official platform documentation. Accessed October 20, 2026.
More from Guides
- How to prompt Luma Uni 1 and Uni 1 Max for images and editsLuma Uni 1 makes and edits images from plain descriptions: say what you want, never what you don't, and name the light. On LUVI it costs 89 credits an image, 96 for an edit; Uni 1 Max costs 220 and 227. Luma suggests 80 to 250 words for a new image and 30 to 100 for an edit.
- How to split an image into layers with Seedream 5.0: what comes back and how to choose the elementsSeedream 5.0 Pro and Flash turn one image into a base image plus up to 16 transparent PNG layers. Leave the prompt empty to separate every main element, or name the ones you want, even by box coordinates. The estimate shows up to 612 credits on Flash; the charge follows the layers returned.
- How to prompt Nano Banana 2.1: prose in, settings out, and no transparent backgroundNano Banana 2.1 reads plain prose and nothing else: describe what should be in the frame, never what shouldn't, and set size and shape in the controls. On LUVI it starts at 80 credits for text-to-image, 83 for an edit and 400 for its video-to-still mode, all at 1k.