PromptMake
Article14 min readImage & describe

GPT Image 2.6 Prompt Patterns: What To Change From 2.5

GPT Image 2.6 prompt patterns: what to change from 2.5 for ChatGPT Images create and edit workflows, GIF loops, markup, and Flare vs Sunburst.

Turn any photo into an AI prompt

No sign-up required. Works with Midjourney, FLUX, DALL-E.

Try Image to Prompt

GPT Image 2.6 is the search label for ChatGPT Images 2.6, OpenAI's native image stack announced 4 November 2026. This page owns the 2.6 delta only: what changed from 2.5 and how I'd retune prompts you already trust. You'll leave with paste-ready create/edit habits, GIF and markup lines, plus Flare vs Sunburst choices for API work. I've kept claims tied to OpenAI's announce post, Help Center, and image-prompting docs. For the 2.5 milestone patterns, use the sibling gpt-image-2-5-prompt-patterns. PromptMake writes prompt text only and doesn't render GPT Image pixels. Soft path: https://promptmake.net/image when a locked still needs Midjourney or FLUX dialect.

What gpt image 2.6 changed from 2.5

OpenAI's consumer product name is ChatGPT Images 2.6 (also Images 2.6 in the announce). The API release paired two model IDs: gpt-image-2.6-flare and gpt-image-2.6-sunburst. If you're typing "gpt image 2.6" into search, you usually want the chat experience first, then the API field list when you're wiring generations or edits. The 2.5 sibling on this site covers what 2.5 changed from prior GPT Image. Stay here for the next step: speed, edit consistency, GIF motion, saved references from the same Help Center wave, plus API complexity and markup that change how I budget words.

OpenAI's 4 November 2026 post markets Images 2.6 as about 50% faster than before, with stronger preservation of characters, products, brand marks, and labels across edit rounds. Help Center echoes the same story: ChatGPT Images 2.6 generates faster and keeps details more consistent from one edit to the next. Treat the 50% figure as OpenAI's comparison, not a bench you ran. Availability is stated as global in ChatGPT and Codex across all subscription tiers on desktop, mobile, web, and related Codex surfaces, with Flare and Sunburst in the API. Check OpenAI's pricing page for current API rates; I'm skipping dollar tables here.

Product surfaces that matter to your prompt text: GIF creation is rolling out in ChatGPT (ask for a looping GIF, or upload a still and ask to animate it; you can't upload an animated GIF for editing). Saved reference photos you reuse with @Name appear in Help Center and release notes for this wave. Templates, Sketch on mobile, selection edits, and prompt sharing continue from the 2.5-era kit. On the API side, official docs add a complexity parameter (low, medium, high; default high when omitted), markup-guided edits with a transparent PNG overlay, SVG and GIF output formats on generations, and a note to omit input_fidelity on GPT-Image-2.6 edits.

This article isn't the general gpt image prompts anatomy guide. It also isn't the 2.5 version-delta page. Those siblings own baseline layers and the prior upgrade. Here we stay on what to change when your client already shows Images 2.6 or when you're migrating an API workflow from 2.5 Flare/Sunburst to 2.6 Flare/Sunburst.

How to adjust gpt image 2.5 habits for gpt image 2.6

Keep the habits that already worked on 2.5: role-labeled references, "change only X" edits, one visual axis per turn, quoted in-frame text, and plain-English aspect ratios. The 2.6 delta is about leaning harder into those clauses because OpenAI positions the model as better at honoring them under faster iteration. If you used to pad prompts with style adjectives hoping the next turn would fix identity drift, cut that padding. Spend the words on locks and motion rules instead. I've watched teams paste a 2.5 "cinematic masterpiece" opener into 2.6 and wonder why the face still wandered while the backdrop got prettier. The upgrade rewards precision, not praise tokens.

Create vs edit still matches Help Center: create describes a new image; edit starts from a ChatGPT output or an upload and describes the change. What shifts is how much you can trust a short preserve list when you swap only a background or a prop. OpenAI's image-prompting guide for the 2.6 line still says define the result, describe visible details, put exact text in quotes, separate changes from constraints, assign roles to references, and iterate one change at a time. That guide is the spine. Your 2.5 notes survive; you'll retune emphasis and add GIF / markup / complexity lines where the client or API exposes them.

Migration checklist from OpenAI's own 2.6 prompting docs: save a 2.5 baseline (prompts, references, hard edits, exact text, product geometry). Test the matching 2.6 model first (2.5 Flare -> 2.6 Flare; 2.5 Sunburst -> 2.6 Sunburst). Keep prompt, references, size, quality, and output format identical on the first compare. Leave complexity at the default for that pass. Inspect identity and labels across a short edit chain. Then tune quality or complexity one knob at a time. Roll out by workflow, not by hope.

Create prompts to retune for gpt image 2.6

Lead with result type, subject, visible facts, then constraints. Skeleton: "Create a [use case] of [subject] in [place]. [Light and materials]. [Medium]. [Composition and aspect]. [Quoted text if any]. [Constraints]." On 2.6, I'd add a one-line identity lock even on first creates when a saved @reference or upload is in play: "@Pepper owns face likeness and skin tone only; invent the scene around that lock." Help Center's example pattern (watercolor of @Pepper) still works; restate what @Pepper controls so a style still doesn't steal the face.

Because OpenAI markets faster generation, resist the urge to stack five scene asks in the first message just because waiting hurts less. A clean first create still beats a kitchen-sink prompt that fails on text and wardrobe at once. Name one medium: photographic, flat vector, watercolor, or 3D product render. Ask for transparent backgrounds in prose when you need them, and verify alpha in the file you download. Exact lettering still wants quoted copy, placement, and "no extra text."

Template and Sketch starts barely change from 2.5. Fill poster facts (audience, headline in quotes, brand colors, blank zones). Let @Sketch own layout while prose owns materials: "Turn this @Sketch into a matte product poster; keep roof pitch from the sketch; no captions." After the first render, drop into the same one-axis edit loop you used on 2.5.

Edit and multi-turn prompts for gpt image 2.6

Write Change first (one noun phrase). Add Preserve for identity and geometry. Name lighting and crop if they must hold. Close with Exclude for logos, watermarks, stray hands, and extra props. Example: "Change only the backdrop to cool gray studio paper. Keep bottle silhouette, label typography, cap, reflections, framing, and light direction. Don't add objects or text." Pair the selection tool or comments when the client shows them, and still name the region in words because Help Center warns highlights aren't always precise.

API users get markup: a transparent PNG overlay matching the first input's pixel size, with the change described in prompt ("Recolor the vase inside the red circle... Remove the red circle from the finished photo"). Omit mask when you send markup. Additional images after the first act as references. Official guidance recommends low complexity for most targeted edits, with medium or high when the scene has many constraints. Omit input_fidelity on GPT-Image-2.6 edit calls; the docs say it isn't supported the way older models used it.

Plan three to five turns: strong first frame, one to three single-axis edits, then constraint cleanup. Restate locks when something drifts. "Same style as before" is weaker than "keep face likeness, navy blazer, window light, and crop." If an old subject from a prior job leaks in, open a new chat. Faster models don't fix polluted threads.

gpt image 2.6 Flare vs Sunburst for chatgpt images API work

Chat users rarely pick Flare or Sunburst by ID; the chat product shows ChatGPT Images 2.6. API and some developer surfaces do. OpenAI's prompting guide calls Flare the smaller model when response time matters most, and Sunburst the base model when detail or precise edits matter more. Both support generation, editing, transparent backgrounds, and the same prompt fundamentals. Flare is also called out in the announce for stronger layout quality on posters and product imagery. Sunburst is positioned for higher-quality generation and greater consistency across edits when quality is set to medium or high. When quality is low, Flare can generate in as little as about six seconds per OpenAI's announce; I'd treat that as their published claim.

Quality and complexity are separate knobs. Quality controls visual detail (texture, small lettering): auto, low, medium, high, xhigh, or max in the 2.6 docs. Complexity controls how much the model plans before rendering: low, medium, or high, with high as the omit-default for both Flare and Sunburst. OpenAI recommends low complexity for most editing tasks, paired with the quality your workflow needs. For open-ended layouts, try higher complexity and compare while holding quality fixed. For fully specified designs, higher complexity may change little. Don't turn both knobs on the same retry if you want a readable A/B.

Output format choices expand the prompt surface: png, jpeg, webp for stills; generations also support svg and gif. SVG fits scalable logos and marks; ask for transparent backgrounds when the mark must sit on arbitrary surfaces. GIF wants motion language (start pose, return pose, camera still or moving, what loops). Size can be auto or documented pixel sizes; custom sizes in the docs want multiples of 16 and aspect ratios roughly between 1:3 and 3:1. Match size in the API fields; also say the aspect in prose when teammates only see the prompt text.

When Flare fits gpt image 2.6 workflows

Start with gpt-image-2.6-flare when your 2.5 Flare (or GPT-Image-2) workflow already met quality bars and you mainly want lower latency. Re-run the same prompts, references, dimensions, and quality first. Use Flare for interactive loops, poster drafts, product stills where layout speed beats micro-texture, and any queue where six-second-class responses (at low quality, per OpenAI) matter more than max fidelity. If Flare fails a hard text or identity check, jump to Sunburst with the same prompt before you rewrite the brief.

Prompt habit for Flare: keep constraints tight and visible. Don't rely on the model inventing a clever composition when you already know the grid. State columns, margins, headline text, and logo clearance once. Low complexity plus a fully specified layout is a common pairing for quick poster edits.

When Sunburst fits gpt image 2.6 workflows

Start with gpt-image-2.6-sunburst when your 2.5 Sunburst path needed maximum detail, or when Flare misses small lettering, product geometry, or multi-turn identity. Use medium or high quality when consistency across edits is the job. Sunburst is the safer first pick for brand packs, face-critical campaigns, and long edit chains where a drifted label costs a reshoot day.

Prompt habit for Sunburst: keep the same preserve lists you wrote for Flare. Don't dump extra adjectives just because the model is the "quality" SKU. Add complexity only when the layout is under-specified and you want the model to plan. After Sunburst passes, OpenAI's migrate notes say try Flare with identical inputs to see if you can reclaim speed without losing the acceptance bar.

New gpt image 2.6 surfaces: GIF, markup, and complexity

GIF prompts need motion grammar, not mood boards. Help Center: ask for a looping GIF and describe what should move, or upload a still and ask ChatGPT to animate it. Adjust playback speed in the GIF viewer after generation. You can't edit an animated GIF you upload; describe edits on GIFs ChatGPT created. Official API prompting examples lock a start pose, a short action, a return to start, and a still camera when you want a clean loop. Example shape: "Make a looping GIF of [subject] [one action]. Start and end in the same pose. Camera still. No new props. No on-image text." Keep motion short and literal. Sprite-sheet style prompts (exact frame grids) appear in OpenAI's announce demos; use them when you need controllable frames, not when you only want a social loop.

Markup prompts name the mark and the change. Draw or upload a transparent overlay that circles or arrows the target. Prompt: say what changes inside the mark, what stays outside it, and that the mark itself must disappear from the final. Chat already had selection and comments; API markup is the portable version of that habit. Don't combine markup with mask in one request when the docs say to omit mask.

Complexity prompts are mostly API fields, but chat users still feel the effect when they under-specify. If you leave trophy design, jersey crests, and banner placement open, higher complexity can rearrange the scene (OpenAI's football paper-cut demo shows that). If you've already fixed every label, keep complexity low and stop blaming the prompt for "overthinking." Quality up when micro type fails; complexity up when composition planning fails. I've wasted retries mixing those two complaints into one rewrite.

Saved references (@Name) belong in the first two sentences of create prompts: name the lock, then invent the scene. Templates still want slot answers, not vibe paragraphs. Sketch still wants layout from the doodle and finish from prose. None of that is rumor; Help Center documents create, edit, GIF, save-reference, templates, Sketch, and the editor's selection caveats as of the Images 2.6 Help update.

Common gpt image 2.6 mistakes after upgrading from 2.5

Mistake 1: Treating 2.6 as a free pass to write vaguer prompts. Faster iteration doesn't replace locks. Keep change-only language.

Mistake 2: Migrating API calls without dropping input_fidelity. Official 2.6 prompting notes say to omit it on edits.

Mistake 3: Asking for a GIF, then uploading an animated GIF to "fix" it. ChatGPT can't edit uploaded animated GIFs per Help Center.

Mistake 4: Turning quality and complexity in the same retry, then guessing which knob fixed the poster. Change one setting at a time.

Mistake 5: Pasting Midjourney flags into ChatGPT Images. --ar and --sref still belong elsewhere; write aspect in prose or use the picker.

Mistake 6: Skipping role labels on multi-image uploads because "2.6 is better at references." Better models still need "image 1 is face only."

Mistake 7: Cloning the entire 2.5 sibling article into team docs. Point people at gpt-image-2-5-prompt-patterns for the prior delta; keep this page for 2.6-only adjustments.

Mistake 8: Inventing GA dates or price sheets from memory. Ship date for Images 2.6 is 4 November 2026 per OpenAI; rates live on their pricing page.

When to export gpt image 2.6 looks with PromptMake /image

Stay in ChatGPT while you explore creates, edits, GIFs, Sketch starts, and templates. That's where gpt image 2.6 already owns the loop for me. Export when you need Midjourney v7 parameters, FLUX material dialect, SDXL negatives, batch runs, or prompt text calibrated from a still you can't describe cleanly.

PromptMake /image drafts model-ready text from a reference photo for Midjourney, FLUX, DALL·E-labeled targets, Stable Diffusion, or Leonardo. It doesn't generate ChatGPT pixels and it doesn't call OpenAI's Images API for you. Soft start: https://promptmake.net/image. As of late 2026, guests get about three image-tool runs per day and registered free accounts about five, with quotas separate from /text.

Pair Recreate Exactly when a ChatGPT Images 2.6 still is the hero composition. Pair Change Style when pose stays and medium changes. Pair Adjust Lighting when chat light grammar was the gap. Pair Create Variation for sibling moodboard briefs. Edit the draft before you paste into another generator; dialect translation is still a human check.

FAQ about gpt image 2.6

What is gpt image 2.6?

GPT Image 2.6 is the common search name for ChatGPT Images 2.6, OpenAI's native ChatGPT image model announced on 4 November 2026. The API side shipped gpt-image-2.6-flare and gpt-image-2.6-sunburst. In chat you'll create from text or templates, edit uploads, use Sketch where available, save references with @Name, and ask for looping GIFs as that feature rolls out. Match the Images label your client shows that day.

How is gpt image 2.6 different from gpt image 2.5?

OpenAI positions Images 2.6 as faster (about 50% in their announce vs prior) and stronger at keeping character, product, brand, and label details across edits. The same wave highlights GIF creation, API complexity controls, markup in the API, and SVG/GIF outputs. Prompt anatomy stays conversational. For the 2.5-from-prior delta, read gpt-image-2-5-prompt-patterns. For baseline gpt image prompts layers without a version angle, read gpt-image-prompts-guide.

How do I write gpt image prompts for gpt image 2.6 edits?

Upload or select the still, say this is an edit, assign roles, list one primary change, and write a preserve list. Example: "Edit this photo. Keep face likeness and hair. Change only the background to soft gray paper. Keep crop and light direction. No text." Use selection or markup for a region, and still name that region in words. On API edits, I'd prefer low complexity for single-detail changes unless the scene is crowded with constraints.

Do chatgpt images mean gpt image 2.6 in my client?

ChatGPT Images is the product family name. GPT Image 2.6 is the November 2026 generation inside that family. Help Center copy as of mid-November 2026 describes ChatGPT Images 2.6 for faster generation and edit consistency. Read the badge in your UI. Create vs edit prompt patterns travel across nearby version numbers; don't freeze docs on a string your account no longer shows.

Can I use Midjourney flags with gpt image 2.6?

No. Flags like --ar, --style raw, and --sref are Midjourney syntax. In GPT Image 2.6 write aspect and style in plain English, or use ChatGPT's aspect ratio picker. Translate to Midjourney only when you export. PromptMake /image helps when a locked still needs Midjourney or FLUX wording without hand-rewriting every clause.

How should I prompt GIFs in gpt image 2.6?

Ask for a looping GIF and name the motion in concrete verbs. Lock start and end poses when you want a clean loop. Keep the camera still unless you want a move. Upload a still to animate if that's the job. Don't upload an animated GIF expecting ChatGPT to edit it. In the API, set output_format to gif and use documented frame fields when you need timing control; check current docs for frame_count and frame_duration_ms limits on your model.

When should I use PromptMake /image after gpt image 2.6?

Use it when a ChatGPT Images 2.6 still is good enough to be a reference and you need export text for Midjourney, FLUX, SDXL, Leonardo, or similar. Open https://promptmake.net/image, pick a goal mode, edit the draft, paste into the other tool. PromptMake doesn't render GPT Image output. Free-tier daily caps on PromptMake are per tool path and sit separate from ChatGPT's own image limits.

Ready to generate your own prompts?

Free. No sign-up required. Works with all major AI models.

Related articles