GPT Image 2.5: What's New, Prompts, and Examples

GPT Image 2.5 brings faster generation and more controlled editing. OpenAI announced ChatGPT Images 2.5 on September 8, 2026, alongside two API models: Flare and Sunburst. The practical choice is speed versus editing precision, rather than a single replacement model for every job. OpenAI's announcement describes the release and rollout.

This guide covers the changes, how to choose a variant, and seven copyable prompts with original images: a poster, a product photo, a background edit, an editorial portrait, a fashion photograph, a comic, and an infographic. Research and examples are dated September 10, 2026.

What changed in GPT Image 2.5?

OpenAI reports improvements to reference fidelity, local edits, and consistency across repeated revisions. Its headline speed claim is up to 50% lower generation latency than Images 2.0; treat that as a vendor claim, not a timing result from this article.

• Reference fidelity: better retention of distinctive subjects, with more natural lighting and textures.
• Focused edits: better adherence to requests that change one element while keeping its surroundings.
• Longer edit sequences: improved retention of earlier changes.
• ChatGPT controls: Sketch, templates, image comments, and shareable prompts. These are product features, separate from choosing an API model.

These are OpenAI's reported changes. Our examples below show individual results, not a benchmark against GPT Image 2.

Flare vs. Sunburst: which should you try?

Start with Flare for quick iterations; test Sunburst when precise edits are the deciding factor. The official model pages describe Flare as the fast everyday option and Sunburst as the option for demanding image generation and editing.

The API IDs are gpt-image-2.5-flare and gpt-image-2.5-sunburst. Both model pages list auto, low, medium, high, xhigh, and max quality settings. Those settings belong to the API request; typing “max quality” into a chat is not equivalent to selecting that parameter.

Our suggested test is a real deliverable with a short acceptance list. For the bottle below: three correct label lines, an intact pump, an unchanged silhouette, and enough room for a headline. Save your accepted source image before comparing edits. A handsome image with the wrong package is still the wrong result.

A practical GPT Image 2.5 prompt structure

Write the brief as deliverable, subject, layout, appearance, exact copy, and constraints. OpenAI's image prompting guide recommends describing the intended image and separating requested changes from details that must remain. The template below is our working version for marketing assets.

Deliverable: [poster, product photo, social graphic]
Subject: [what the viewer should notice first]
Layout: [shape, subject position, space for text]
Appearance: [materials, palette, light, medium]
Exact copy: "[approved wording]"
Constraints: [must-have details and unwanted additions]
For an edit: change [one thing]; preserve [specific details].

For example, “luxury skincare ad” leaves the package, copy, lighting, and composition undecided. “Amber pump bottle on limestone, cream label, window light from upper left, empty left third” makes those decisions reviewable. If the result fails, you can name the missed requirement.

Give each block one job. Put the wording in its own sentence. Specify whether you want flat artwork or a photograph of a printed object. If you need a real product, supply its photograph as the reference instead of expecting a text description to reconstruct it.

Example 1: a poster with exact event copy

This ceramics poster uses a small amount of copy and two clear regions: typography on the left, a jug on the right. This is the exact prompt used:

Create a polished landscape 3:2 promotional poster for a fictional ceramics market. Editorial Swiss-inspired graphic design, warm ivory paper with subtle grain, deep cobalt-blue typography, small vermilion accents. Exact text, no other words: "FORM & FIRE" as the largest headline in the upper-left; "CERAMICS MARKET" below it; "12 OCTOBER" and "10 AM - 6 PM" as clearly separated small lines at lower-left. A beautifully photographed sculptural cobalt-blue ceramic jug with a curved handle occupies the right half, casting a soft realistic shadow. Keep generous margins and a clear gap between type and object. Flat front-facing finished artwork, not a photograph of a poster on a wall. No logos, no watermark.

Full size

What the output did: it included all four requested text strings without additional wording. The headline dominates, the time and date stay together, and the jug occupies the right half. The glaze has visible variation instead of a perfectly uniform blue surface.

What we would revise: the prompt asked for small vermilion accents, but the output includes a prominent red circle behind the jug. That is a useful example of an instruction that leaves room for interpretation. For a quieter version, we would ask: “Remove the red circle behind the jug. Keep the small red rule above the date, and preserve every other element.” That follow-up is a suggested revision, not another tested output.

How to adapt it: replace the four copy strings with approved event details and swap the jug for your subject. Keep the spatial instructions. Avoid adding a paragraph of sponsor copy to this layout without giving the typography more room.

Example 2: a product photo with room for an ad

The product prompt reserves the left third for later copy and uses a fictional bottle so the exercise can be judged without pretending to reproduce a real SKU. The label, glass, and stone each have a different requested material treatment.

Create a landscape 3:2 studio product photograph of one fictional skincare bottle. An amber glass cylindrical bottle with a black pump, front-facing cream paper label, centered slightly right on a pale warm limestone plinth. Exact label text on three lines: "MORA", "DAILY WASH", "250 ml". Soft window light from upper left, believable amber translucency, fine paper texture and a gentle contact shadow. Muted sand-colored seamless background. Leave the left third open for future advertising copy, but add no text outside the bottle label. The bottle must be fully visible, with no props, extra products, people, logos beyond the fictional label, or watermark. Premium but natural photography, not a glossy 3D render.

Full size

What the output did: the three label lines are readable, the pump is inside the frame, and the bottle sits to the right of open background. Paper grain, glass highlights, and the rough stone distinguish the materials. The scene also includes soft background shadows, which the brief did not specify individually.

How to adapt it: for your own product, attach the approved packshot and replace the invented bottle description with a preservation instruction. Name the cap, package proportions, label artwork, and product color. For more composition ideas, see our AI product photography guide and product lifestyle scene tool.

Example 3: change the background, keep the product

Use the previous result as the edit input and request one visible change. Here, we supplied the original bottle image and asked for a sage background:

Edit the supplied product photograph. Change only the sand-colored background to a muted sage-green background. Preserve the amber bottle, black pump, cream label, all label text ("MORA", "DAILY WASH", "250 ml"), limestone plinth, camera position, crop, object scale, original lighting direction and shadows. Do not add leaves, props, graphics or new text. Keep the same landscape dimensions. The bottle and plinth must not be recolored.

Full size

What the output did: the background became sage green, the label wording remained readable, and the bottle and plinth stayed in the same general composition. It added no leaves, props, or extra copy.

What still changed: fine texture on the stone and subtle lighting details differ. This is visually similar preservation, not a pixel-identical background replacement. The result is a useful concept variant; an exact catalog master needs closer inspection. OpenAI also notes that repeated edits can change protected details and recommends compositing when a region must remain pixel-identical. See the editing guidance.

How to adapt it: use a color from your campaign palette and keep the product constraints explicit. Inspect the first change before requesting another. If the label drifts, return to the approved source rather than building the next variant on the damaged one.

Example 4: an editorial portrait with natural texture

The header portrait uses a fictional adult woman, window light, and an open wall to create a magazine-style image with room for a headline. The prompt describes visible surfaces and framing rather than relying only on “photorealistic.”

Create a landscape 3:2 editorial portrait for a creative technology magazine. An attractive adult woman in her late twenties with warm brown eyes, light freckles, and loose dark wavy hair, wearing a cobalt-blue silk blouse and small gold earrings. Waist-up, seated at a sunlit artists studio table, body angled slightly away while looking directly into the camera with a relaxed confident expression. Position her on the right half; leave the left third as softly lit warm ivory plaster wall with subtle light and shadow, suitable for a headline added later. A few softly blurred paintbrushes at the far right suggest a working studio. Natural side-window light, fine skin texture, individual flyaway hairs, visible silk weave, restrained film grain, believable photographic color. Face and both shoulders completely inside the frame with generous headroom. No written text, no graphics, no logos, no beauty-filter smoothing.

What the output did: the woman occupies the right side, with a softly lit ivory wall on the left. Freckles, loose strands of hair, and folds in the blue blouse are visible. The gold earrings and studio brushes carry through. The model added a hand-under-chin pose, which the prompt left open.

What to inspect: there is less headroom than requested, and the blouse reads more like a matte woven fabric than lustrous silk. If either matters, revise that detail explicitly. At full size, check the hand against the jaw, the earring edges, and the hairline before treating a portrait as finished.

How to adapt it: describe your location, wardrobe, gaze, and the space needed for copy. To depict a particular person, attach an approved reference; this text-only example creates a fictional face and does not demonstrate likeness preservation.

Example 5: full-body fashion and movement

A full-body fashion image needs instructions for the feet, garment, pose, and surrounding space. Here we asked for a walking pose with a small bag and enough margin to keep both sandals inside the frame.

Create a portrait 2:3 full-body fashion editorial photograph. An attractive adult East Asian woman in her late twenties with a short black bob walks across a sunlit terracotta courtyard, wearing an oversized ivory linen pantsuit, a rust-colored silk scarf tied at the neck, and simple dark brown leather flat sandals. Show her entire body, both feet and their contact shadows, with comfortable space above her head and below her sandals. Capture a natural mid-step pose, one hand loosely holding a small structured tan leather bag, the other arm swinging naturally. One model only. Architectural arches form a quiet repeating background. Late afternoon directional sunlight, realistic linen creases, subtle fabric movement, warm terracotta and ivory palette. High-end location fashion photography, natural skin texture, no floating fabric, no text, no logos, no collage.

Full size

What the output did: both feet fit inside the composition, one hand carries the bag, and the forward step reads clearly. The linen has visible creases, the scarf stays at the neck, and the courtyard arches provide depth without introducing another person.

What to inspect: the fabric looks cleanly pressed in some areas and heavily wrinkled in others. The scarf hangs mostly still despite the walking pose. Decide whether that suits your brief before requesting more movement. Inspect fingers, bag handles, trouser hems, and sandal straps at full size.

How to adapt it: keep “entire body” and the explicit framing margins when footwear matters. For an actual garment campaign, supply clothing references and judge the seams, closures, and fit against them. A convincing invented outfit is not proof that a real garment was reproduced accurately.

Example 6: a three-panel illustrated story

A comic prompt needs a sequence of actions and a short description of the features that should repeat. This example assigns one story beat to each panel and repeats the woman's outfit and the cat's markings as continuity requirements.

Create a landscape 3:2 illustrated comic strip with exactly three equal vertical panels in a single horizontal row, separated by clean ivory gutters. Hand-painted gouache and colored-pencil texture, warm storybook editorial illustration, muted teal, apricot and cream palette. The same attractive adult woman appears in every panel: chin-length curly auburn hair, round gold glasses, teal overalls over a cream T-shirt. Her companion is one small orange cat with a white chest and white front paws. Panel 1, left: she sits at a desk sketching in an open notebook while the cat peers over the desk edge. Panel 2, middle: the cat lies across the notebook; she pauses with pencil held above the page and gives it an amused sideways look. Panel 3, right: she has placed a small empty cardboard box beside the notebook; the cat curls up in the box while she resumes sketching. Preserve her hair, glasses, outfit, and the cat markings across all three panels. Clear readable actions and expressive faces, no speech bubbles, no captions, no lettering anywhere. Keep every scene within its own panel.

Full size

What the output did: it produced exactly three panels with clear gutters and no speech bubbles or lettering. The character's hair, glasses, and overalls remain recognizable, and the cat moves through all three requested actions. Paper-like texture and visible strokes give the image an illustrated finish.

What we would revise: the box in the last panel overlaps the notebook rather than sitting neatly beside it. The cat's white markings also vary in shape. For a tighter result, ask to move only the box to a clear area of desk and use an approved character reference when those markings need to match.

How to adapt it: keep the panel count explicit and write each action separately. Introduce only the characters and props needed for the scene. This is a single generated strip; it does not establish character consistency across separate requests or a whole illustrated book.

Example 7: an infographic with a defined reading order

Give an infographic its content before asking for visual treatment. We supplied four steps, their order, the exact captions, and the connections between them. The graphic illustrates a proposed creative workflow, not statistics or a factual dataset.

Create a landscape 3:2 editorial infographic titled exactly "FROM BRIEF TO CAMPAIGN". Warm ivory background, crisp navy sans-serif type, small cobalt and coral accents, restrained paper-cut illustration with subtle tactile shadows. Exactly four equal-width numbered columns from left to right, with one arrow between each neighboring column, all arrows pointing right. Column 1: number "01", heading "BRIEF", icon of a notepad, one caption "Define the goal". Column 2: number "02", heading "CREATE", icon of an image frame containing a simple sun and hills, one caption "Explore directions". Column 3: number "03", heading "REVIEW", icon of a magnifying glass over a check mark, one caption "Check the details". Column 4: number "04", heading "EXPORT", icon of a download tray, one caption "Prepare the files". Use only that exact text, each string once. Title spans the top with generous margin, numbers and headings align across the four columns, captions share a baseline. No charts, percentages, extra steps, logos, decorative gibberish, or watermarks. Finished flat graphic, not a photographed sheet.

Full size

What the output did: the title, four numbered headings, and four captions match the supplied wording. Three arrows connect neighboring columns in the requested direction. The icons follow the notepad, image, review, and download brief, with consistent paper-cut styling.

What to inspect: the headline nearly fills the width, leaving little space for a longer title. The captions are readable here but need a separate check at your intended display size. For a real chart, compare every number, label, and graphical relationship with the source data; this example tests layout and copy only.

How to adapt it: replace the workflow with your own approved steps and maintain one caption per column. If you need more explanation, add it in the surrounding article or slide notes rather than compressing a paragraph into each icon.

What these examples do and do not establish

We generated six original images and one source-based edit with GPT Image 2.5 through Codex's built-in image tool on September 10. All seven are first outputs for the prompts shown. Landscape images retain their original 1536 × 1024 dimensions; the fashion photograph retains its original 1024 × 1536 portrait dimensions. There was no image-generation API call, model comparison, or retouching after generation.

These examples illustrate GPT Image 2.5 prompting, without comparing Flare against Sunburst. For separate examples from both API variants, OpenAI's official prompting guide includes labeled outputs. We have kept our prompts and results together so you can evaluate what each instruction produced.

Use the brief in a repeatable creative workflow

Keep the prompt, source image, and accepted output together. For a campaign, record which details are approved before asking for alternate backgrounds or formats. Our recommended order is composition first, copy second, then a small set of variants from the accepted image.

Dream Pixel Forge's AI ad generator is the relevant workflow for turning a campaign brief into creative. As of this article's source check, Dream Pixel Forge's GPT image route still uses GPT Image 2. This article does not announce a GPT Image 2.5 integration.

Access, pricing, and transparent backgrounds

OpenAI announced rollout across ChatGPT, ChatGPT Work, and Codex, including all tiers, plus availability of both API variants. Check the release page for the rollout details.

The Flare and Sunburst model pages list the same token rates: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens. These are not fixed per-image prices. OpenAI explicitly says the GPT Image 2 calculator does not estimate 2.5 token consumption, so do not carry an old per-image estimate into a new budget.

Both 2.5 variants support transparent backgrounds. For API use, request background="transparent" and PNG or WebP output, then check the actual alpha channel. A checkerboard painted into an image is not transparency. This capability is documented in the current prompting guide; our bottle examples use opaque backgrounds.

Start with one asset you can judge closely. The poster and infographic expose copy and hierarchy; the portraits expose framing and texture; the comic exposes continuity; the product edit exposes unwanted changes. Those checks will tell you more about fit for your work than an impressive image with no brief attached.