Run GPT Image 2.6 Sunburst, the detail-focused version of OpenAI's November 2026 image model, right in your browser. Write a prompt to generate stills, looping GIFs, or SVG vectors, or upload up to four references and describe the one change you want. The panel shows the credit cost of every run before you click Generate.
It is the image model OpenAI released on November 4, 2026. Inside ChatGPT it ships as ChatGPT Images 2.6, and OpenAI says it generates images 50% faster than the previous release while keeping characters, products, and brand details steadier from one edit to the next. Developers get two versions. Flare is tuned for speed and quick drafts, while Sunburst trades some speed for higher detail and stronger consistency at medium and high quality. This page runs the Sunburst endpoints for both text to image and editing, with the new GIF, SVG, and reasoning controls exposed in the panel above. Saved reference libraries and frame-by-frame GIF editing remain ChatGPT features, so on this page each run carries its own uploads.
Sunburst is the higher-detail build of the model. Fabric weave, product labels, skin texture, and reflections hold up when you zoom in, and five quality tiers, from Low to Max, let you decide how much detail each draft is worth. Low is enough to judge a composition; Max earns its cost on hero shots and large prints.
Name one change, such as a new jacket color or a swapped background, and the framing, face, and lighting stay where they were. OpenAI built Sunburst for consistency across rounds, so a fifth revision still matches the first approved shot. When a detail must not move, say so in the prompt as well.
Headlines, packaging panels, and product grids come out with more orderly spacing than in earlier releases. Put the exact wording in quotes, say where it sits, and proof the small print before you publish, since tiny lettering can still slip. It often helps to settle the layout first and fix wording in a separate edit.
Reasoning effort controls how much planning happens before rendering. OpenAI recommends Low for most edits and Medium or High for dense scenes, infographics, or prompts with many constraints. Higher effort takes longer, but on this page it does not change the credit cost. Low stays the quickest setting for rapid iterations.
The work that gains most from this model is work that goes through review. A brief turns into drafts, a draft collects notes, and each note should cost one edit instead of a fresh start. That loop is where Sunburst's edit consistency pays off. The four jobs below are where teams lean on GPT 2.6 image generation today. Most start in text to image to explore directions at Low quality, switch to the edit tab once a direction is approved, and save High or Max for the final pass. Because every setting and its credit cost are visible before you generate, drafts stay cheap and you only pay for detail once the brief is settled.

Every option below sits in the generator panel, and the button shows the exact credit cost before you run. GPT Image 2.6 pricing follows output size, quality tier, image count, GIF frames, and reference uploads, because the gpt-image-2.6 Sunburst endpoint is billed by token rather than at a flat rate. Some settings depend on each other. SVG exists only in text to image, transparency needs PNG or WebP, and a GIF shares one pixel budget across its frames. The panel resolves these conflicts as you change options, so it never sends a combination the model would reject. When an option is greyed out, the current format or size cannot support it.
Pick GIF as the output format, then set 2 to 16 frames and a per-frame duration from 20 to 2,000 ms. Frames share one pixel budget, so a 1024×768 loop allows up to 10 frames, a 1024×1024 square allows 7, and the largest sizes are unavailable for GIFs.
The edit tab accepts up to four references, such as a product, a face, a style board, and a logo. Output size snaps to the supported ratio closest to your upload, and each reference adds a credit or two to the run. Name each upload's role in the prompt, for example "use image 2 for the logo".
Quality runs Low, Medium, High, Extra high, and Max, while reasoning effort runs Low, Medium, and High. Quality drives detail and cost. Reasoning drives planning time and leaves the price unchanged, so raise it only for crowded or rule-heavy scenes. Extra high and Max are new tiers that earlier GPT Image versions here do not offer.
Auto lets the model decide, Opaque returns a finished frame, and Transparent returns a cutout with a real alpha channel. Transparency works with PNG and WebP only, because JPEG cannot store alpha and would flatten the cutout onto a solid fill. Pick Opaque when the image needs a finished backdrop for a post or banner.
Text to image can return an SVG file instead of pixels, which suits icons, badges, stickers, and simple logos. Describe flat shapes and a short color palette in the prompt, since vectors work best for clean graphics rather than photographic scenes. Each SVG run returns a single file.
Choose 1:1, 4:3, 3:4, 3:2, 2:3, 16:9, or 9:16, with squares up to 2880×2880 and widescreen up to 3840×2160 in either orientation. Generate up to four stills per run in PNG, JPEG, or WebP, with 0–100 compression for JPEG and WebP. Bigger sizes and higher tiers cost more credits, so preview small and scale up the keeper.
Answers on versions, pricing, GIF and SVG output, references, precise edits, and how this page compares with ChatGPT Images 2.6, plus tips for getting a clean first result.
It suits images that go through review: ad creatives, product photos, posters, packaging comps, icons, and short animated loops. Sunburst keeps details steady across successive edits, so it pays off most when an image needs several rounds of notes rather than one lucky generation.
Yes, it is one release under two names. OpenAI calls it Images 2.6 inside ChatGPT and Codex, and GPT Image 2.6 in the API. Searches also use ChatGPT Image 2.6, GPT Images 2.6, and the no-space spelling gpt image2.6, and all of them point to this model.
Version 2.6 adds looping GIF output, SVG export, adjustable reasoning effort, and two quality tiers above High. OpenAI also reports 50% faster generation and steadier consistency across edits. Version 2.5 is still a good pick for plain stills, with a fixed price for each quality tier.
This page runs gpt-image-2.6 Sunburst for both text to image and editing. Flare is OpenAI's speed-first version and can return a low-quality image in about six seconds. Sunburst takes longer but gives finer detail and stronger consistency across edits, which matters more for work you publish.
Cost follows your settings. A Low-quality 1024×1024 image is 3 credits, Medium is 5, High is 6, and Max is 22. Larger sizes cost more, up to 72 credits for a 2880×2880 image at Max. Each reference adds about one or two credits, and the button always shows the exact price first.
There is no separate free mode, because every run spends credits from your account. The cheapest test is a single Low-quality square at 3 credits, so try your prompt there, check the composition and wording, then raise quality only for the version you plan to keep.
Choose GIF as the output format, pick a size that leaves room for frames, and set the frame count and duration. Describe one subject and one simple motion, such as rising steam or a blinking eye, and keep the camera fixed so the last frame flows back into the first.
Upload the image in the edit tab, then write two things: what should change and what must stay identical. For a change limited to one region, open the image editor, choose Local adjustment, paint over that area, and describe the edit. Everything outside your selection stays untouched.
No. The GPT Image 2.6 API is meant for developers building the model into their own apps, which takes a provider account and code. Here you sign in, type a prompt or upload references, and download results in the browser, with no keys or SDKs to manage.
The model is aimed at commercial-style output such as ads, product imagery, and social posts. Usage rights still depend on the provider terms and on what you upload, so make sure you own or license any reference photos, logos, or faces, and check the applicable terms before publishing client work.
Looking for prompt ideas and model comparisons? Read the model guides blog.