GPT Image 2.5 Sunburst guide — prompts, text and clean layouts

GPT Image 2.5 Sunburst is OpenAI's flagship image model, and on eroq it is the one you pick when the picture has a job description: a headline that must be spelled right, a product that must sit in the left third, a layout that has to match the sketch on your whiteboard. It follows prompts closely, keeps composition clean, and it is the lineup's pick for sharp in-image text. This guide is for designers, marketers and anyone who writes briefs: how to prompt it, how to get lettering that reads, and where it stops being the right tool.

The numbers

• Price: 31 credits / image at every format, about $0.31 at the entry pack, where a credit is roughly a cent.
• Speed: typically about 15 seconds a take (a planning estimate, not a guarantee).
• Formats: 1:1, 3:4, 4:3, 9:16 and 16:9, each rendered natively.
• Takes: 1 to 4 per run, billed and refunded one by one. A full run of four is 124 credits.
• Inputs: a text prompt. On eroq there are no reference pictures and no Edit mode for this model.
• Plan: any plan: a pay-as-you-go credit pack is enough, no subscription.
• Content: moderated upstream by its provider. A refused prompt returns content_blocked and costs nothing.
• API id: gpt-image-2-5-sunburst on POST /v1/images/generations.

Write the prompt like a design brief

Sunburst rewards the order a good brief already follows. Five parts, in this sequence:

1. What the artifact is. "A product photograph", "a flat-color event poster", "a mobile onboarding screen". Naming the artifact sets a hundred defaults at once: margins, framing, how much of the frame the subject takes.
2. The subject and where it sits. "A ceramic mug in the lower right third", "the climber small against the cliff, top left". Position words are cheap and Sunburst listens to them.
3. The background. Seamless paper, a blurred kitchen, a solid color. Name the color properly: "deep teal" beats "dark".
4. The light. One sentence is enough: "soft window light from the left, a gentle shadow falling right".
5. The words. The exact text in quotes, where it goes and what it is set in.

Then choose the ratio in the Format control instead of describing it. "A vertical poster" plus 3:4 is redundant but harmless; "a vertical poster" rendered in 16:9 is a fight you will lose.

Keep one clear idea per image. Two competing subjects split the composition. If a frame needs two products, make one the hero and give the other a job: "a smaller matching lid leaning against its base".

Text that actually reads

This is why most people open Sunburst. A few habits make the difference between a clean headline and a near-miss:

• Quote the exact words, in the case you want: the headline "OPEN STUDIO WEEKEND" in capitals.
• Keep it short. A headline, a tagline, a label or a button. The fewer words you ask for, the cleaner each one comes out. Body copy belongs in your layout tool, not in the render.
• Say where and how. "Set in a bold condensed sans across the top", "a small serif caption under the photo", "centered on the label".
• Give the text a surface. A poster, a label, a sign, a screen, a T-shirt. Words floating in a scene have nothing to anchor to.
• Check at full size, then re-roll. There is no Edit mode on Sunburst, so a misspelled take is a re-render, not a touch-up. Rendering four takes at once is the efficient way to find the clean one.

For the cross-model view, see the guide to legible AI text and the Sunburst vs Flux 3 typography showdown.

Composition you can steer

Words that move the layout reliably: "centered and symmetrical", "rule of thirds, subject on the left", "generous negative space on the right for copy", "flat lay shot from directly above", "isometric view", "three-quarter view", "tight crop", "wide establishing shot". For structured layouts, describe the structure before the content: "three illustrated steps in a row, each with a numbered circle and a one-word caption". For UI mockups, list the screen elements in reading order, top to bottom, with each label quoted.

Example prompts

1. Event poster, 3:4. Artifact first, one subject, two text blocks with placement and type style.

   A flat-color screen-print poster for a weekend pottery market: a stack of glazed bowls in terracotta and cream at the center, a sun shape behind them, generous margins, the headline "CLAY DAYS" in a bold rounded sans at the top and "SAT–SUN · RIVER HALL" in small capitals at the bottom, warm off-white paper texture.

2. Product hero, 4:3. Placement plus empty space you can fill with real copy later, and only a one-word logo in the render.

   A clean studio product photograph of a sage-green insulated water bottle standing on a pale stone block in the right third of the frame, soft key light from the left, a long gentle shadow, empty space on the left, the printed logo on the bottle reads "TRAILMATE" in white.

3. App onboarding screen, 9:16. Elements in reading order, every string quoted, a single button.

   A mobile app onboarding screen for a plant-care app shown flat on a white background: a large illustration of a monstera leaf at the top, the heading "Water less. Grow more." below it, one rounded green button labeled "Get started" near the bottom, clean modern sans-serif, lots of white space.

4. Infographic-style banner, 16:9. The structure is stated before the content, and each label is one word.

   An infographic-style banner explaining pour-over coffee: three illustrated steps in a row, a grinder, a pour-over cone and a cup, each with a numbered circle and a one-word caption, "GRIND", "BLOOM", "POUR", flat vector style, cream background, dark brown and copper palette.

5. Desk flat lay, 1:1. No text at all: a reminder that the layout control is worth having on its own.

   An overhead flat lay of a designer's desk: a sketchbook open to pencil thumbnails, a steel ruler, three markers in coral, teal and mustard, a phone showing a color palette, everything aligned at right angles, soft daylight, calm and orderly.

6. Shop sign, 4:3. Lettering on a physical surface in a photographic scene.

   A small bakery storefront in the morning, a hand-painted wooden sign above the window reading "BREAD & BUTTER", fresh loaves visible behind the glass, a bicycle leaning by the door, soft overcast light, documentary photograph, 35mm.

When Sunburst is not the pick

• More atmosphere than precision: its sibling GPT Image 2.5 Flare costs the same and renders bolder. Sunburst vs Flare splits the difference.
• A final where every detail must hold: Nano Banana Pro, at 124 credits / image.
• A natural-light photograph: Flux 3 Image.
• Your real product, logo or face: Sunburst cannot see your files. Compose the layout here, then rebuild it on Eroq One, which reads up to ten reference pictures.

For a ready-made library, 25+ GPT Image 2.5 prompts covers posters, products, layouts and UI. Or open the image studio, pick Sunburst and start with two takes of the poster prompt above.

FAQ

How do I get GPT Image 2.5 Sunburst to spell text correctly?

Put the exact words in quotes, keep them to a headline or a label, say where they sit and in what type style, and give them a surface such as a poster or a sign. Render several takes and keep the cleanest; a misspelled take is a re-render, since there is no Edit mode.

How long does GPT Image 2.5 Sunburst take to render?

Typically about 15 seconds a take. It is a planning estimate: a busy moment can add queue time, and the studio shows how long each take actually took.

Can GPT Image 2.5 Sunburst use my logo or a photo of my product?

Not on eroq: it is text-to-image only and reads no reference pictures. Describe the layout on Sunburst, then use Eroq One, which builds a shot from up to ten reference pictures, when the real logo or product has to appear.

What aspect ratios does GPT Image 2.5 Sunburst support?

Five: 1:1, 3:4, 4:3, 9:16 and 16:9, each rendered natively and at the same price. 3:4 suits posters, 9:16 suits phone screens and stories, 16:9 suits banners and slides.