You open a GPT Image 2.6 tool, type a prompt, and before you can hit generate there's a toggle: Flare or Sunburst. No tooltip explains it. Pick wrong and you either wait longer than you needed to on a throwaway draft, or you get a draft-grade image on a job that needed a final.
OpenAI released Images 2.6 on November 4, 2026, with two API models, gpt-image-2.6-flare and gpt-image-2.6-sunburst. Most coverage so far stops at "Flare is fast, Sunburst is quality." That's true, but it doesn't tell you which one to use for a poster, a GIF, a product shot built from twelve reference photos, or the fifth round of edits on a character. It also doesn't answer the first question most people have: where can you actually use each mode? This guide covers both, using OpenAI's own documentation (as of November 5, 2026). Where something hasn't been published, the article says so.
Quick Answer: Which Mode for Which Job
| Your job |
Pick |
Why |
| Exploring ideas, 4+ variations |
Flare |
OpenAI built it for fast iteration |
| Social posts, quick mockups |
Flare |
OpenAI says Flare improved layout for posters and product visuals |
| Final hero image, print, client delivery |
Sunburst |
OpenAI's "detailed visual work" model |
| Multi-round edits on one character or product |
Sunburst |
OpenAI's one specific claim: more consistency across edits at medium/high |
| Product edit with many reference images |
Sunburst |
The more an edit must preserve, the more consistency matters |
| Short GIF drafts |
Flare |
Motion ideas need several tries; faster rounds help |
| Text-heavy poster |
Flare to draft, Sunburst to finish |
Work out the layout fast, then lock the final |
| SVG logo or icon |
Either, lean Sunburst |
A logo is usually a keeper, and geometry has to stay exact |
The short rule: Flare until you know what you want, Sunburst once you're keeping it.
Where Can You Use GPT Image 2.6 Flare and Sunburst?
There are three practical routes, and they differ in how much control you get over the mode.
| Where |
How you get Flare / Sunburst |
Good for |
| ChatGPT and Codex |
Images 2.6 runs inside the chat; no documented mode picker |
Casual use inside a conversation, ChatGPT's own editing tools |
| OpenAI API |
Set the model ID: gpt-image-2.6-flare or gpt-image-2.6-sunburst |
Developers building pipelines and apps |
| This site, in the browser |
A Flare / Sunburst toggle on one tool page |
Picking the mode yourself without writing code |
ChatGPT and Codex. OpenAI rolled Images 2.6 into ChatGPT and Codex at launch. As of November 5, 2026, OpenAI hasn't documented a way to choose Flare or Sunburst there; the product decides. If you mainly want to know which mode produced your image, or to force one, this route doesn't give you that.
The OpenAI API. Here you pick the model explicitly by its ID and get every API option, including the complexity setting, markup-guided edits, SVG and GIF output. You need an API key, an OpenAI developer account and some code. For reference, OpenAI lists the same token rates for both models: $5 per 1M text input tokens, $8 per 1M image input tokens and $30 per 1M image output tokens. OpenAI's guide adds that token counts "can differ by model and quality setting," so the same request can cost differently on each model. Check the usage field on real responses.
This site. The GPT Image 2.6 generator puts both modes behind one toggle in the browser, with no API key and no ChatGPT subscription. What it covers:
- Text to Image and Image to Image (edits), with up to 16 reference images per edit (JPG, PNG or WEBP, up to 10MB each)
- PNG, GIF and SVG output
- 1K, 2K and 4K resolution, low / medium / high quality, and up to four PNGs per run
- Aspect ratios from 9:21 to 21:9, plus "auto" in edit mode to keep the input's shape
What it doesn't cover: markup-guided edits, a complexity control (it uses the API default, high) and per-frame GIF editing. If you need those, use the API or ChatGPT.
The practical advantage is the side-by-side test. Because both modes share one page, you can run a prompt on Flare, flip the toggle and run the identical prompt on Sunburst, then compare the two results yourself instead of trusting anyone's benchmark.
What OpenAI Actually Says About Each Mode
The two names aren't marketing tiers on one model. OpenAI ships them as two separate model IDs, each with its own documentation page, and the wording is deliberate:
- GPT-Image-2.6 Flare: "Image generation and editing for fast iteration." The model page recommends it for "interactive image generation and editing workflows."
- GPT-Image-2.6 Sunburst: "Image generation and editing for detailed visual work," recommended "for detailed compositions and image editing workflows."
The launch announcement adds three specifics:
- Images 2.6 generates images 50% faster than before (the previous generation, 2.5).
- With quality set to low, Flare can generate an image in as little as six seconds.
- Sunburst brings higher-quality generation and greater consistency across edits when quality is set to medium or high. OpenAI also says Flare improves layout quality for posters, product imagery and other professional visuals.
The ChatGPT Images 2.6 system card summarizes it the same way: Flare is optimized for faster generation. Both models share every new 2.6 feature in the API.
For 2.5, OpenAI's prompting guide called Flare "the small model, optimized for speed" and Sunburst "the base model, optimized for quality." The 2.6 positioning carries over one-for-one (see the GPT Image 2.5 complete guide).
Head-to-Head: Flare vs Sunburst
| Dimension |
GPT Image 2.6 Flare |
GPT Image 2.6 Sunburst |
| OpenAI positioning |
Fast iteration, interactive workflows |
Detailed visual work, detailed compositions |
| Published speed claim |
As little as ~6 s at low quality |
None published (shares the 50%-faster claim) |
| Edit consistency claim |
Not singled out |
Greater consistency across edits at medium/high |
| Layout claim |
Improved layout for posters and product imagery |
Higher-quality generation overall |
| Best moment in a project |
Early: exploring, testing, drafting |
Late: finals, delivery, long edit chains |
| API token rate |
$5 text in / $8 image in / $30 image out per 1M |
Identical |
| GIF, SVG, markup edits (API) |
Supported |
Supported |
| Batch API |
Not supported |
Not supported |
Speed: What the "6 Seconds" Claim Covers
The six-second figure is the only absolute latency number OpenAI published for 2.6, and it comes with conditions. It applies to Flare, at low quality, and says "as little as," which describes the fast end of a range, not an average.
What the claim doesn't cover:
- Medium and high quality, the settings most people use for anything they keep.
- Sunburst, which gets only the general "50% faster" claim.
- Edits with reference images, higher complexity (OpenAI says it "can increase latency") or 4K output.
So the honest conclusion is a direction: Flare is the lower-latency model, clearest at low quality. That matters most while exploring. If you need ten attempts to find a composition, a faster model keeps you in a quick loop. Once you've found the image, seconds on one final render matter much less than getting it right.
If latency matters, measure it yourself with prompt, references, size and quality held identical, as OpenAI's prompting guide recommends. On this site that means same prompt, same settings, flip the toggle.
Quality and Consistency Across Edits
This is the difference that should drive most decisions, and it's narrower than people assume.
OpenAI didn't say Sunburst makes every image better. It said Sunburst brings higher-quality generation and greater image consistency across edits at medium or high quality. Consistency across edits means the parts of an image you didn't ask to change stay the same when you run the next edit: a character's face, a product's label, a logo's geometry, the exact shade of a brand color.
That matters most over several rounds. Each edit is a new generation conditioned on the previous output, so small drift (a slightly different eye shape, a softened letter in a label) compounds. Every drift you catch is a round spent repairing instead of improving, so a model that drifts less saves more the longer the chain gets.
Two practical consequences:
- Don't judge Sunburst on low quality. OpenAI's consistency claim is specifically for medium and high. Comparing the modes at low quality tests a setting where Sunburst's headline advantage isn't promised.
- Single, simple edits are less of a differentiator. A one-shot recolor or text swap is a short chain. Third-party testing of the 2.5 pair (OmniaKey, November 1, 2026) found both models handled single recolors and chalk-text changes; the gap showed up on a harder remove-and-add instruction. That's one small test on the previous generation, but it fits OpenAI's framing: the more an edit has to preserve, the more Sunburst is worth it.
What about Flare's quality? Don't read "fast" as "bad." OpenAI specifically calls out Flare's improved layout for posters and product imagery. For a lot of social and marketing work, Flare at medium or high is the finished product.
As of November 5, 2026, Arena-style leaderboards list 2.6 results only as preliminary, so treat early rankings as early.
Flare and Sunburst share one tool page here. Run the same prompt in both and switch with one click, no API key needed.
Open GPT Image 2.6 →
Scenario by Scenario
Drafts and exploration
Flare, low or medium quality, 1K. This is the work Flare was built for. Try composition, framing and color ideas quickly, generate up to four PNGs per run, keep the one that works, and only then think about Sunburst. Expect several rounds here; that's the point of a fast model.
Finals and hero images
Sunburst, high quality, at the resolution you'll deliver. If the asset is going on a landing page, in print or to a client, it's the one image people will look at closest. That's where Sunburst's detail is worth a longer render. Render at the final size rather than upscaling later, so fine detail is generated, not stretched.
Multi-round edits on one subject
Sunburst, medium or high. This is the use case OpenAI's main claim is about. Say you're refining a character over five edit rounds: changing the outfit, then the pose, then the background, then the lighting, then the expression. On a model that drifts, you may spend two of those rounds just restoring a face that changed. On a model that holds the face, all five rounds move the image forward. The longer the edit chain, the better Sunburst looks.
Product edits with up to 16 reference images
Sunburst. This site's Image to Image mode accepts up to 16 reference images. A heavily referenced edit usually means a specific product must look exactly right: the label, the cap shape, the logo placement. That's the preservation job Sunburst is positioned for. Use Flare first if you're only testing scene ideas around the product, then switch to Sunburst with the same references once the scene is settled.
GIFs
Flare to find the motion, then decide. Motion ideas usually take several tries: the loop doesn't close, the speed feels wrong, the subject drifts between frames. On this site, GIFs come in 4, 8 or 12 frames at fast (100 ms), medium (200 ms) or slow (400 ms) per frame, in 1:1, 16:9, 9:16, 4:3 or 3:4.
Frame count affects detail. More frames means each frame gets fewer pixels: on this site a square 4-frame GIF renders each frame at 1024×1024, an 8-frame GIF at 1008×1008, and a 12-frame GIF at 816×816. So for a detailed animated subject (a mascot with a readable logo, for example), a 4- or 8-frame Sunburst GIF keeps details steadier than a 12-frame version. For a quick reaction loop, 12 frames on Flare is usually enough. There's more detail in the GPT Image 2.6 GIF guide.
ChatGPT's GIF editor lets you adjust individual frames. This site doesn't offer per-frame editing, so to change a GIF you edit the prompt and regenerate.
Text-heavy posters and infographics
Draft on Flare, finish on Sunburst. OpenAI singled out Flare's improved layout for posters, so it's a good choice for working out where the headline, subhead and fine print go. Once the layout works, move to Sunburst at high quality for the final, where small text and consistent letterforms matter most. A 32,000-character prompt limit leaves plenty of room for exact copy. Put every word you want rendered in quotes.
SVG logos and icons
Lean Sunburst. A logo is almost never a throwaway draft, and its geometry has to stay exact. SVG on this site is available from Text to Image only, one file per request, with no background option. On the API side, OpenAI also documents SVG for generation only, not for the edit endpoint. To vectorize an existing raster logo, use the image to SVG converter.
If You Need X, Choose Y
- If you need a lot of variations quickly, choose Flare at low or medium.
- If you need the image to survive five edits without the face changing, choose Sunburst at medium or high.
- If you're uploading a stack of product photos, choose Sunburst once the scene idea is set.
- If you need a GIF concept by lunch, choose Flare; re-render the winning idea on Sunburst only if frame detail matters.
- If you're on the API and cost-sensitive, test both and compare
usage. The token rate is equal, but token counts may not be.
- If neither mode suits your style, try the same prompt on Nano Banana Pro or Seedream 5.0 Pro, or go back to the GPT Image 2.5 generator to compare generations.
The "Draft on Flare, Finish on Sunburst" Workflow
This workflow fits both modes' strengths, and you can run all of it on one page by switching modes on the same prompt. Here it is for a 2K product poster:
- Explore on Flare. In Text to Image, choose Flare, 2K, low quality, 4 outputs. You're judging layout and mood, not detail.
- Pick and tighten. Choose the best layout. Rewrite the prompt with exact text in quotes and any color or placement fixes. Run Flare again at medium quality, 2 outputs.
- Finish on Sunburst. Switch to Image to Image, upload the chosen draft (and the product photo if you have one), flip the toggle to Sunburst, 2K high, 1 output.
You get fast loops while the idea is still moving, and Sunburst's consistency once the idea is fixed. Running every exploration on Sunburst at high quality would make each early round slower without helping you decide anything.
Two tips that keep the workflow working:
- Keep the aspect ratio fixed from step 1. Switching from 4:5 to 16:9 at the finish line means a new composition, not a refinement. This site offers 1:1, 9:21, 9:16, 2:3, 3:4, 4:5, 5:4, 4:3, 3:2, 16:9 and 21:9.
- Stay on Sunburst once you switch. If the final needs three more changes, make them on Sunburst. Switching back to Flare mid-chain gives up the consistency you switched for.
How This Compares to the 2.5 Split
GPT Image 2.5 introduced the Flare/Sunburst pair in September. With 2.6, both names moved up a version with the same jobs.
| Area |
2.5 Flare / Sunburst |
2.6 Flare / Sunburst |
| Speed |
Up to 50% lower latency vs GPT Image 2 |
50% faster again, Flare as little as ~6 s at low |
| Edit consistency |
Sunburst positioned for edit precision |
Sunburst: greater consistency across edits at medium/high |
| Complexity control (API) |
No |
low / medium / high, default high |
| Output formats (API) |
Raster images |
Adds SVG and animated GIF |
| Token rates |
$5 / $8 / $30 per 1M |
Unchanged |
OpenAI's migration advice: for an existing 2.5 Sunburst workflow, try 2.6 Sunburst first, and the same for Flare. A "Flare drafts, Sunburst finals" routine carries over unchanged. The GPT Image 2.6 complete guide covers the rest of the release.
Run the Flare-to-Sunburst workflow on your own prompt
Draft on Flare, flip to Sunburst for the final, all on one page. PNG, GIF and SVG at up to 4K, with no API key or ChatGPT subscription.
Try GPT Image 2.6 →
Compare with GPT Image 2.5 →
FAQ
What is the difference between GPT Image 2.6 Flare and Sunburst?
Flare is OpenAI's fast-iteration model: it can produce a low-quality image in as little as about six seconds. Sunburst is the detailed-work model, with higher-quality generation and greater consistency across edits at medium or high quality. Both support the same 2.6 features in the API, including GIF, SVG, markup edits and the complexity setting.
Where can I use GPT Image 2.6 Flare and Sunburst?
In ChatGPT and Codex, Images 2.6 runs inside the chat, but OpenAI hasn't documented a mode picker. In the OpenAI API, you choose by model ID (gpt-image-2.6-flare or gpt-image-2.6-sunburst) with an API key. On this site, both modes sit behind one toggle in the browser, for text-to-image and edits with up to 16 references, with no API key needed.
Which GPT Image 2.6 model is better for editing?
For edits that must preserve a face, product or logo across several rounds, choose Sunburst at medium or high quality. That's the setting OpenAI's consistency claim covers. For a single simple change like a recolor, Flare is often enough, and it's faster.
How fast is GPT Image 2.6 Flare?
OpenAI says Flare can generate an image in as little as six seconds with quality set to low. That's a best case for one setting. OpenAI hasn't published timings for medium or high quality, for Sunburst, or for edits with many reference images.
Do Flare and Sunburst cost the same on the OpenAI API?
They have the same token rates ($5 text input, $8 image input, $30 image output per 1M tokens), matching GPT Image 2.5. OpenAI notes token counts can differ by model and quality, so the cost per image can differ. Compare the usage field on real requests.
Can I choose Flare or Sunburst inside ChatGPT?
OpenAI hasn't documented a model selector in ChatGPT as of November 5, 2026. In the API you set the model ID directly. On this site, both modes appear as a toggle on the GPT Image 2.6 tool page.
What does the complexity setting do in GPT Image 2.6?
It's an API setting that controls how much the model works out the layout before rendering: low, medium or high, with high as the default for both models. OpenAI suggests low for focused edits and higher levels for dense text and complex layouts, and notes that higher complexity can increase latency. This site uses the default high setting.
Sources
- Introducing Images 2.6 (OpenAI) — release date, 50% faster claim, Flare ~6 s at low quality, Sunburst edit consistency at medium/high, Flare layout improvements.
- GPT-Image-2.6 Flare model page — "fast iteration" positioning, token rates, supported endpoints, no Batch support.
- GPT-Image-2.6 Sunburst model page — "detailed visual work" positioning, token rates, supported endpoints.
- OpenAI image generation guide — complexity setting and defaults, SVG and GIF output, token counts differing by model and quality.
- OpenAI image prompting guide — 2.5 Flare as the small speed model and Sunburst as the base quality model, comparison method, migration advice.
- OpenAI API pricing — current image token rates.
- ChatGPT Images 2.6 system card — Flare optimized for faster generation; both models' editing and detail-preservation improvements.
- OpenAI Developers launch post on X — Sunburst keeps details more consistent from one edit to the next at medium or high quality.
