GPT Image 2.6 Prompts: 9 Prompts We Ran on each::labs

Most prompt lists are written by people who never pressed generate. You can tell. The prompts are long and pretty, and nothing in the post says what came back.

These GPT Image 2.6 prompts are different for one reason: we ran every one of them on each::labs this morning, on GPT Image 2.6 Sunburst and GPT Image 2.6 Flare, and we'll tell you what each one returned, which settings we used, and the one sentence that made the difference.

How to Read These GPT Image 2.6 Prompts

Every prompt below comes with three settings, because with this model the settings are half the prompt. The model (Sunburst for detailed work, Flare for fast iteration, as OpenAI's image generation guide splits them), the complexity (low, medium or high), and the size. We kept quality at medium for every run, so the differences you read about come from the prompt and the complexity, not from rendering more pixels.

Complexity is the setting people misread. OpenAI describes it as how much the model plans before rendering, separate from quality, and it defaults to high if you leave it out. Its guide is specific about when to move it: lower complexity "can suit focused edits, such as changing a background or a sign", while higher complexity "can help with dense text, complex layouts, and relationships between objects". It also warns that high "can increase latency and introduce unwanted changes". Keep that warning in mind. Two of our results below show it in action.

If you want the full model breakdown first, our GPT Image 2.6 test on each::labs covers Sunburst against Flare. This post is the copy-and-paste companion.

Prompt Structure That Works: Scene, Subject, Details, Then Rules

OpenAI's prompting guide for earlier GPT Image models recommends an order: the scene, then the subject, then the key details, then the constraints. Put literal text in quotes. For edits, say what changes and list what stays. All of it held up on 2.6.

What we'd add after today: end with a rule, not an adjective. "No other text." "No logos." "Equal size and spacing." The closing constraint is the line GPT Image 2.6 enforced most reliably, and we regretted leaving it out both times we did.

Text and Layout Prompts

Text is where GPT Image models earn their reputation, so we started there. Every run spelled every word correctly. The differences were in what else the model decided to add.

1. A gig poster with exact type

A minimalist A3 gig poster for a jazz night. Background: deep teal paper texture. Top third: the headline "LATE SET" in tall condensed cream letters. Middle: a single brass trumpet drawn in loose ink lines, angled 30 degrees. Bottom left, small: "Friday 21:00". Bottom right, small: "Harbor Room". No other text.

We ran it twice: Sunburst at high complexity and Flare at low, both 1024x1536. Both placed all three lines of text exactly where we asked and added nothing else. The interesting difference was the trumpet. Sunburst read "drawn in loose ink lines" literally and drew it as cream line art on the teal. Flare rendered a golden, colored trumpet. Same words, two readings of "brass". If the illustration style matters, say the color as well as the medium.

2. An infographic with numbered steps

A clean flat infographic on an off-white background titled "How Cold Brew Works". Four numbered steps left to right, each a simple icon with a one-line caption under it: 1 "Grind coarse", 2 "Add cold water", 3 "Steep 16 hours", 4 "Filter and serve". Thin arrows connect the steps. Palette: espresso brown, cream, one accent of orange. Sans-serif type, generous spacing.

Sunburst, high complexity, 1536x1024. Title, four numbered steps, four captions, arrows, palette: all there, all spelled right, with an orange underline under the title as a small design flourish. It also wrote "16 h" on the steeping jar, which is the kind of helpful extra that's fine on an infographic and annoying on a product shot. This is exactly the dense-layout job OpenAI says high complexity is for.

3. A chalkboard menu

A chalkboard menu hanging on a brick wall in a small bakery, photographed straight on. Heading at the top in hand-lettered chalk: "TODAY". Below it, three lines, each an item with a dotted leader to the right: "Cardamom bun", "Rye sourdough", "Pear galette". A small chalk drawing of a wheat stalk in the bottom right corner. Warm morning light from the left. No other text.

Sunburst, high complexity, 1024x1536. The heading, the three items, the dotted leaders and the wheat stalk all landed, in convincing chalk lettering, with the light coming through a window on the left. High complexity also dressed the set: a basket of bread and a small vase of flowers appeared on the sill. Lovely for a mood shot. If you need the board alone, say "the board fills the frame".

4. A magazine cover, and the line we forgot

A print magazine cover. Masthead across the top: "FIELD NOTES". Cover photo: a woman in a yellow raincoat crouching beside a tide pool on a grey rocky coast, overcast light. Lower left, a cover line in bold white: "The Winter Coast Issue". Under it, smaller: "12 walks worth the wind". Barcode in the bottom right corner. Clean editorial layout. No other text.

Flare, high complexity, 1024x1536. That last sentence is in the version above because we didn't have it in the run. Every line we asked for came out perfect, and Flare also invented an issue line under the masthead, with a volume number and a date, plus digits under the barcode. Add "No other text" to every layout prompt and you remove the most common reason a good image gets bounced.

Product and Brand Prompts

Product work rewards precise camera language and a clear note about where the copy will go.

5. A product hero with room for copy

Studio product photo of a matte black ceramic pour-over coffee dripper on a pale oak counter. Soft window light from the left, a thin curl of steam, three coffee beans in the foreground slightly out of focus. Shot on an 85mm lens at eye level, shallow depth of field, clean negative space on the right for copy. No text, no logos.

Flare, medium complexity, 1536x1024. It came back in about 9 seconds end to end, among the quickest runs of the day, and it followed the brief closely: light from the left, steam, three beans soft in the foreground, an empty right half for a headline. One quirk worth knowing: the dripper came back with a mug handle on it. "Pour-over dripper" is a less common object than "mug", and the model leaned toward the familiar shape. For unusual products, describe the form ("a cone-shaped dripper on a flat ring base, no handle"), or better, edit from a real product photo.

6. A die-cut sticker with a transparent background

A die-cut sticker of a cheerful cartoon avocado wearing tiny round sunglasses, thick white outline around the whole shape, flat colors, transparent background.

Flare, low complexity, 1024x1024, with background set to transparent and output format PNG. Ask for transparency in both places: in the prompt and in the request. The file came back with a real alpha channel, the sticker sitting on empty space and ready to drop onto a product page. The each::labs pages support transparent backgrounds for PNG and WebP output.

7. An icon set that stays consistent

A set of six flat line icons for a travel app, arranged in a 3 by 2 grid on a white background: a suitcase, a plane, a map pin, a passport, a camera and a tent. Uniform 2px navy stroke, rounded corners, no fill, equal size and spacing, no text.

Flare, low complexity, 1536x1024. Six icons, one stroke weight, one color, one grid, no stray labels. Low complexity was plenty, because the hard part (consistency across a set) is stated as a rule, not left to taste. OpenAI's API also offers SVG output for GPT Image 2.6, but the each::labs pages return PNG, JPEG or WebP today, so treat this as a raster mockup to hand to a designer or trace.

Character Prompts for Turnarounds and Mascots

8. A three-view character sheet

A character turnaround sheet for a small explorer robot with a round copper body, one large teal camera eye and short tank treads. Three views side by side on a plain light grey background: front, side, back. Same proportions, same colors and same scuffs in every view. Clean 3D render, soft studio light, no text.

Sunburst, medium complexity, 1536x1024. Three views, one robot. The copper body, the teal eye, the treads and the scuffs matched from view to view, which is the whole point of a turnaround. Sunburst also gave the robot accessories we didn't ask for: an antenna, a small dish and a rolled pack. Because they were consistent across all three views, they read as design choices rather than mistakes. If your character is already designed, list every part and end with "nothing else".

Edit Prompts: Change One Thing, Name Everything Else

For edits we used GPT Image 2.6 Flare Edit on the coffee dripper image from prompt 5.

9. Swap a surface, keep the product

Change only the counter surface to white Carrara marble with soft grey veins. Keep the black dripper, the paper filter, the steam, the three coffee beans, the plant, the window light and the framing exactly as they are.

Flare Edit, low complexity, 1536x1024. The oak became marble. Everything on the keep list stayed put: the dripper (handle and all), the filter, the curl of steam, the three beans in the same spot, the plant in the window. That's the edit OpenAI describes when it says low complexity suits focused changes like a background or a sign, and our run agrees.

The habit that made it work is the keep list. Name every element you care about, not just "everything else". A list gives the model something concrete to protect. For a harder target, use GPT Image 2.6 Sunburst Edit, and point at a one-of-a-kind detail ("the mug with the label") rather than a position that may be ambiguous.

Five Rules From Ten Runs

Pull the results together and a short rulebook falls out.

Quote every word that must appear, and end layout prompts with "No other text". The two layouts that picked up extra words, the magazine issue line and the "16 h" on the infographic jar, were the two prompts without that line.

Use high complexity for layouts with many parts, and expect it to add set dressing. Bread baskets, flowers, an orange underline. If the frame must stay empty, say so.

Use low complexity for focused edits and simple, rule-bound sets like icons. It was fast and it was exact.

Describe unusual objects by shape, not only by name. The model will reach for the nearest familiar object when the name is rare.

Say the color when you name a medium. "Ink lines" and "brass" pulled our two trumpets in different directions.

Where These Prompts Go Next on each::labs

A good workflow uses both models. Explore a prompt on Flare at low or medium complexity, generating several options at once (the pages allow up to 10 images per request). Move the winner to Sunburst at high complexity for the final. Fix small things with an edit at low complexity. All four GPT Image 2.6 pages sit behind the same each::labs API, so each step is a model id change.

From there, a still can become the first frame of a clip with one of the image to video models, or the whole chain can run as an each::labs flow. If you're coming from the previous version, our GPT Image 2.5 prompting guide still applies, and our GPT Image 2.6 launch post explains the new SVG, GIF and markup features in OpenAI's own API. To compare with other models, browse the text to image catalog.

Frequently Asked Questions

What makes good GPT Image 2.6 prompts?

Order and rules. Describe the scene, then the subject, then the details, put every literal word in quotes, and finish with constraints such as "No other text" or "equal size and spacing". In our runs the closing constraint was the most reliably obeyed line in the prompt.

Should I use Sunburst or Flare for these prompts?

Flare for exploring and for simple, rule-based images like icons, stickers and quick product shots. Sunburst for dense layouts, text-heavy designs and character sheets where many details have to agree. OpenAI positions Sunburst for detailed visual work and Flare for fast iteration.

What complexity setting should I use in GPT Image 2.6?

Low for focused edits and simple sets, high for layouts with many parts. It defaults to high when omitted. OpenAI notes that high can add latency and unwanted changes, and in our runs it did add props nobody asked for, so drop it when the frame must stay clean.

Can GPT Image 2.6 make transparent PNGs on each::labs?

Yes. Set background to transparent with PNG or WebP output, and ask for a transparent background in the prompt too. Our sticker test came back with a real alpha channel.