GPT Image 2.6 Explained: Flare vs Sunburst, GIFs, Pricing & the Hidden Cost

GPT Image 2.6 Explained: Flare vs Sunburst, GIFs, Pricing & the Hidden Cost

On 4 November 2026, OpenAI released GPT Image 2.6, the first update to its image model since version 2.5 arrived in September — and it changes more than the version number suggests. ChatGPT can now turn a prompt or a photo into a looping animated GIF, remember the people, pets and characters you use most, and keep them looking the same from one picture to the next. Developers get two distinct API models to choose from: Flare, built for speed, and Sunburst, built for detail and edit consistency. There is also a new complexity control, image markup in the API, and SVG export for designers who need scalable vector files. This guide explains what GPT Image 2.6 actually does, how Flare and Sunburst differ, what each image really costs — including the pricing trap most launch coverage glossed over — and which model you should pick for your work.

A graphic designer working with a stylus at a desk, the kind of creative workflow GPT Image 2.6 is built for

Table of Contents

GPT Image 2.6 at a Glance

GPT Image 2.6 is OpenAI’s November 2026 image-generation update, and it is the most practical one the company has shipped in a while. The headline features are not abstract benchmark gains. They are workflow fixes: animation, memory and precision. Here is the short version before the detail.

  • Launched: 4 November 2026, in ChatGPT, Codex and the API, on every ChatGPT subscription tier from Free to Pro.
  • Two API models: gpt-image-2.6-flare for fast iteration (as little as six seconds at low quality, according to OpenAI) and gpt-image-2.6-sunburst for detailed work and consistent edits at medium or high quality.
  • GIFs: turn a prompt or an image into a looping animated GIF, adjust its speed and edit individual frames. In the API, animations run from 2 to 64 frames.
  • Saved references: save photos of yourself, a pet or a character once, name them, and reuse them in later conversations with an @ mention — no more re-uploading the same picture every time.
  • New developer controls: a complexity setting (low, medium, high), image markup for targeted regional edits, and SVG export for scalable vector graphics.
  • Speed: OpenAI says generation is about 50% faster than before. Independent timing is thinner — one early measurement put a medium-quality image at 4.84 seconds against 7.08 seconds for version 2.5.
  • The pricing catch: token rates are unchanged at $5 per million text-input tokens, $8 per million image-input tokens and $30 per million image-output tokens — but GPT Image 2.6 spends up to 4.2 times more output tokens at low quality and 3.3 times more at medium, so a low-quality 1024×1024 image now costs about $0.025 instead of about $0.006.
  • Scale: OpenAI says image generation across ChatGPT and the API has now passed 4 billion images per week, up from about 3 billion in September 2026.

What Is GPT Image 2.6?

GPT Image 2.6 is the image model behind ChatGPT Images 2.6, OpenAI’s November 2026 refresh of the image tools inside ChatGPT. If you use ChatGPT, the update reaches you as new abilities in the chat window: GIFs, saved references and faster, more consistent edits. If you build software, it reaches you as two new model identifiers in the Images API, plus four new controls that did not exist in version 2.5.

The release matters because of where it lands. OpenAI’s own figures put weekly image generation at more than 4 billion images across ChatGPT and the API, up from roughly 3 billion in September. That is an enormous audience, and most of those people are not benchmarking image models. They are making product shots, social posts, storyboards, logos and birthday cards. GPT Image 2.6 is aimed squarely at them: the three features a regular user will notice first — animation, character memory and edit consistency — are all about doing ordinary creative work with less friction, not about winning a leaderboard.

That said, the leaderboard story is real too. On launch day, DesignArena reported Sunburst taking first place on its Image Arena leaderboard, and by 7 November Artificial Analysis listed GPT Image 2.6 Sunburst (max) at rank one for text-to-image, with its editing ranks overlapping the leading competitors. Treat those as early signals, not settled verdicts: leaderboards move fast, and the editing ranks in particular were close enough that no single winner was clear in the first days.

Availability is broad but not instant everywhere. OpenAI says GPT Image 2.6 is available in ChatGPT and Codex on every subscription tier, on web, desktop and mobile, and through the API. GIF creation, however, was still rolling out at launch — if the option does not appear in your ChatGPT yet, that is the phased rollout, not a plan restriction. The same applies to Codex, OpenAI’s coding agent, which gets the new image models as part of the same release.

What GPT Image 2.6 is not

It is worth clearing up three misunderstandings that spread quickly after the announcement. First, GPT Image 2.6 is not a video model. GIFs are short loops of still frames — useful for a mascot waving, a product spinning or a storyboard beat, but not a substitute for a video generator. Second, it is not cheaper than version 2.5, despite the identical rate card; the pricing section below explains why the per-image cost actually rose at low and medium quality. Third, it is not a new ChatGPT plan or a paid add-on. In ChatGPT itself there is no per-image price at all: image generation is included in your subscription, and OpenAI publishes no per-plan image or GIF counts. The per-image maths in this article applies to the API, where you pay per token.

Flare vs Sunburst: The Two New Models

The biggest structural change in GPT Image 2.6 is that developers no longer get one image model. They get two, with the same price per token and different jobs. Flare is the fast one. Sunburst is the careful one. Choosing between them is the single most important decision in this release, so here is the comparison in full.

Feature GPT Image 2.6 Flare GPT Image 2.6 Sunburst
Model ID gpt-image-2.6-flare gpt-image-2.6-sunburst
Built for Speed and fast iteration Detail and edit consistency
Fastest generation As little as 6 seconds at low quality (OpenAI) Slower; tuned for quality, not speed
Best quality setting Low to medium Medium to high
Best for Drafts, high-volume jobs, posters, product imagery Final assets, repeated edits, detailed compositions
Token rates $5 text input / $8 image input / $30 image output per 1M tokens Identical rates
Cost per image Same as Sunburst at the same size and quality Same as Flare at the same size and quality
New controls Complexity, markup, GIF and SVG output Complexity, markup, GIF and SVG output

The pricing row deserves emphasis, because it surprises almost everyone: Flare and Sunburst cost the same per token. OpenAI’s own calculator treats them as a single entry, “GPT Image 2.6 (Sunburst and Flare)”. That means the choice between them is a trade of seconds for detail, not dollars. At the same size and quality setting, both models use roughly the same number of output tokens — within a few dozen of each other, or under a tenth of a cent apart. You pick Flare when waiting is the bottleneck and Sunburst when quality is.

OpenAI positions Flare for interactive generation: brainstorming, quick mock-ups, bulk product shots and anything where you will generate ten versions and keep one. Sunburst is positioned for the opposite end of the workflow: the final hero image, the brand asset that must survive five rounds of edits without the character’s face drifting, and the crowded composition with many elements and constraints. Early coverage agrees with that split. One launch-week comparison advised using Flare for drafts, high-volume jobs, posters and product imagery, and Sunburst for final assets and repeated edits at medium or high quality.

There is one more difference worth knowing before you commit to either model in production. The API model pages currently mark the Images endpoints (generations and edits) as supported, but mark the Responses API, Chat Completions, Realtime, Batch and fine-tuning as unsupported for both 2.6 models. If your app integrated version 2.5 through the Responses tool, do not assume that integration carries over unchanged — check the current model pages before you ship.

GIF Creation: ChatGPT Finally Animates Your Images

The feature most ChatGPT users will try first is also the simplest to describe: GPT Image 2.6 can turn a written prompt or an uploaded image into a looping animated GIF. Describe what should move — a mascot waving, a product rotating, a sketch character blinking — and ChatGPT produces an animation you can share. You can adjust the playback speed and edit individual frames after it is made.

The animation styles OpenAI highlights give a sense of the intended audience: cartoon illustrations, chibi characters and 1990s anime sit alongside product shots, mascots and storyboards. That is a creator’s list, not a film-maker’s one. A GIF is a loop, not a video, and the format’s limits shape what it is good for: short, repeatable motion. A logo that bounces once, a character that waves, a “before and after” that flips. Anything that needs a camera move or a narrative arc still belongs in a video tool.

On the technical side, the API documentation is specific. A GIF can have between 2 and 64 frames (9 by default), and each frame can show for 20 to 2,000 milliseconds (200 by default). Those two settings — frame count and frame duration — are the whole animation control surface in the API, and they are enough to build anything from a snappy two-frame flicker to a slow, deliberate sequence. In ChatGPT itself the controls are presented in plain language rather than numbers, which fits the audience: most people animating a mascot do not want to think in milliseconds.

Three limits matter, and all three were true at launch. First, GIF creation in ChatGPT was still rolling out, so some users saw the new image model before they saw the animation option, as TechRepublic’s launch coverage noted. Second, in the API, GIF output sits behind a feature flag that was switched off and undocumented at launch, so third-party apps cannot rely on it yet — it is a ChatGPT-first feature for now. Third, ChatGPT cannot currently edit an animated GIF that you upload. You can make one and you can tweak the one you made, but you cannot hand ChatGPT an existing GIF and ask it to change a frame. If your workflow depends on editing GIFs you already have, that gap is worth knowing before you plan around the feature.

What GIFs are actually good for

The honest answer is narrower than the launch excitement suggests, and more useful for it. GIFs shine in four places: social posts that need motion without the weight of a video file; product pages where a small loop shows a feature better than a static shot; messaging stickers and reactions built from a mascot or a pet; and storyboards, where a rough animated beat communicates timing that a still frame cannot. They are a poor fit for anything longer than a few seconds, anything with dialogue, and anything where smooth, high-frame-rate motion matters. Used inside those limits, they are the most immediately fun part of GPT Image 2.6 — and the part most likely to bring new users into image generation at all.

Saved Reference Images: Your Characters Stay Consistent

If GIFs are the most fun part of GPT Image 2.6, saved reference images are the most useful. The problem they solve is one every regular image-generation user knows: the model does not remember your dog between requests. Keeping the same character looking the same across five scenes used to mean re-uploading the same photo every time and hoping the model held the likeness. Saved references remove that step.

Here is how it works. You upload a picture of yourself, a pet or a character once, save it as a reference and give it a name. In later conversations you retrieve it by typing @ followed by its saved name, and the model uses it as the reference for the new image. The character, product or brand detail stays consistent across edits and across sessions, because the reference is stored rather than re-described in every prompt.

That consistency is the feature’s real value, and it shows up in three common jobs. A small business can keep its product looking identical across a whole catalogue of lifestyle shots. A creator can keep a mascot or a comic character on-model across a series. And anyone making portraits — a headshot series, a family illustration, a profile-picture set — can stop fighting the model’s tendency to redraw the same face differently each time. If you make portraits often, the same consistency problem is why dedicated headshot tools exist at all, and it is worth comparing the approaches in our guide to the best AI headshot generator before you commit to one workflow.

A photographer holding a camera, the kind of reference photo GPT Image 2.6 can save and reuse across images

There is a privacy question attached to this feature that OpenAI’s announcement does not fully answer, and it deserves a plain statement. Saving reference photos of real people makes it easier to place someone into an animated scene that never happened. Coverage of the launch noted that the announcement does not explain how OpenAI protects real people’s likenesses in GIFs, or whether the animations carry a watermark. Until OpenAI says more, the sensible rule is the obvious one: only save reference images of yourself, people who have agreed to it, and characters you own. Do not use someone else’s photo as a saved reference.

Image Markup: Point at Exactly What to Change

The second workflow fix in GPT Image 2.6 is image markup, and it addresses the oldest complaint about AI image editing: “change one thing, change everything.” Anyone who has asked an image model to fix a single detail — a sign’s spelling, a hand, a background object — and watched it redraw the whole picture knows the frustration. Markup is OpenAI’s answer.

In the API, markup lets a developer select the exact region of an image to change, so the edit is applied where it is pointed and the rest of the picture is left alone. The feature already existed inside ChatGPT; GPT Image 2.6 brings it to the API, which is where it matters most for products. An app that lets users tap the part of a photo they want changed can now send that selection to the model as a precise instruction instead of a vague sentence.

Markup pairs naturally with the complexity setting and with Sunburst’s edit consistency. The intended workflow is: mark the region, set a low complexity for a small targeted change, and use Sunburst at medium or high quality so the untouched parts of the image survive the edit intact. For a single quick fix, Flare at low complexity is the faster route. Either way, the point is the same — editing becomes a scalpel instead of a hammer — and for anyone who edits images professionally, that change alone may be worth more than the GIFs.

SVG Export: Vector Graphics from the API

Designers get the most specialised new feature in GPT Image 2.6: SVG export. An SVG is a vector file — the image is stored as instructions (“draw a curve from here to there”) rather than as a grid of pixels — so it stays sharp at any size and a designer can still edit its shapes afterwards. That makes it the right format for logos, icons, illustrations and anything destined for a website, an app or print.

There are real constraints, and they are all in the API documentation. SVG output is an API-only feature for now; you cannot get an SVG from ChatGPT. To request one, you set the output format to “svg” and ask for exactly one image at a time. The SVG that comes back is static vector content only — no scripts, no external resources and no embedded raster images. And the edit endpoint cannot return an SVG, nor can it take one as an input, so you cannot hand an SVG back to the model and ask for a tweak. You generate the vector, then edit it in a design tool.

A designer sketching by hand, the kind of illustration work GPT Image 2.6 SVG export is built to support

Even with those limits, SVG export is a meaningful addition. Until now, an AI image generator handed you pixels, and a logo made of pixels is a logo that blurs the moment it grows. A vector file that a designer can open, recolour and scale is a different kind of output — it slots into a professional workflow instead of sitting outside it. The obvious comparison is with the open-weight image models that designers self-host, and the trade-offs there are covered in our FLUX 3 Image Klein guide. If your work is logos and brand marks, SVG export alone may be the reason to try GPT Image 2.6 first; if your work is photography, it will not matter to you at all.

The Complexity Setting: How Much Planning Before It Draws

Alongside the two models, GPT Image 2.6 adds a control called complexity, with three levels: low, medium and high. It controls how much planning the model does before it renders an image. OpenAI’s guidance is straightforward: use low complexity for quick edits, and medium or high for crowded scenes with many elements and constraints.

The default matters here, because it is not the cautious one. If you do not choose a setting, complexity defaults to high. OpenAI’s own documentation warns that higher complexity can increase latency and introduce unwanted changes — the model plans more, and sometimes that extra planning rearranges things you wanted left alone. The announcement’s recommendation is the practical version of the same advice: low complexity for most editing tasks, saving medium or high for scenes where many elements must hold together.

In practice, complexity is the dial you reach for when an image keeps coming out wrong in a specific way. A simple edit that keeps changing the background? Drop to low. A busy poster where the text, the product and the background keep fighting each other? Raise it to high and accept the slower generation. Combined with markup, it gives developers a genuinely finer level of control than version 2.5 offered: point at the region, set the planning level, pick the model. That is the shape of a maturing API, and it is the part of this release that professional users are most likely to still be using a year from now.

GPT Image 2.6 Pricing: The Rates Did Not Change — Your Bill Did

Here is the part of the GPT Image 2.6 story that most launch coverage got wrong, or skipped entirely: the price. OpenAI’s token rates for the new models are identical to version 2.5 — $5 per million text-input tokens, $8 per million image-input tokens and $30 per million image-output tokens, with cached input cheaper ($1.25 for text, $2 for images). On paper, nothing changed. In practice, a low-quality image now costs about four times what it cost a week ago.

The reason is token usage. GPT Image 2.6 spends far more output tokens per image than version 2.5 did at the same quality setting. Independent measurements taken the day after launch put the increase at up to 4.2 times more output tokens at low quality and 3.3 times more at medium. Because you pay per token, not per image, the unchanged rate card produces a very different bill.

Quality (1024×1024) Output tokens (2.6) Cost per image (2.6) Cost per image (2.5)
Low ~823 ~$0.025 ~$0.006
Medium ~1,449 ~$0.043 ~$0.013
High ~1,756 ~$0.053 ~$0.053
Extra high ~3,100 ~$0.094 —
Max ~7,024 ~$0.21 ~$0.21

Read that table carefully, because it contains the whole pricing story. At high quality and above, nothing changes: a high-quality 1024×1024 image costs about $0.053 on both versions, and a max-quality image costs about $0.21 on both. The increase is concentrated exactly where the cheap work lives — low and medium quality, the settings used for drafts, bulk generation and quick iterations. A thousand low-quality images now costs about $24.69 in output tokens instead of about $6, and a thousand medium-quality images costs about $43.47.

There is a second cost catch, and it matters for anyone running image generation at volume. The Batch API, which gave version 2.5 a half-price rate, is not supported for either 2.6 model. As of the days after launch, the pricing page’s Batch tab listed 2.5 at half price and did not list the 2.6 models at all. If your pipeline depended on batch pricing to keep costs down, GPT Image 2.6 is more expensive twice over: more tokens per image, and no batch discount to soften it.

How much does GPT Image 2.6 cost per image in ChatGPT?

Nothing per image. In ChatGPT itself, image generation is included in your subscription on every plan from Free to Pro, and OpenAI publishes no per-plan image or GIF counts. The dollar figures in this section apply only to the API. That distinction is the reason the pricing trap caught people out: ChatGPT users never see a bill, so the token increase is invisible to them, while API users building products on top of the model feel it on every call. If you are choosing between the two routes, the rule of thumb is simple — occasional use belongs in ChatGPT, where the cost is your subscription; systematic, high-volume use belongs in the API, where you should budget on the new token counts, not the old ones.

Photo editing software on a computer screen, where GPT Image 2.6 pricing per image decides how many edits you can afford

How to keep GPT Image 2.6 costs down

  • Prototype in ChatGPT, produce in the API. Do your prompt iteration where it is free — inside your subscription — and only call the API for the final images you actually need.
  • Use Flare at low quality for drafts. Flare is built for exactly this, and at low quality a draft costs about $0.025. Generate ten drafts, keep one, then spend the $0.053 on a single high-quality Sunburst final.
  • Do not pay for quality you will not use. A thumbnail, a social preview or a test render does not need high quality. The gap between low and high is roughly double the price per image.
  • Watch the batch gap. If you relied on version 2.5 batch pricing, do not migrate a bulk pipeline to 2.6 until Batch support appears — or until you have re-run the numbers without the discount.
  • Set complexity to low for edits. Lower complexity means less planning, which usually means fewer tokens spent on changes you did not ask for.

How Fast Is GPT Image 2.6 Really?

OpenAI says GPT Image 2.6 generates images about 50% faster than before, and that Flare can produce an image in as little as six seconds at low quality. Those are the company’s numbers, and they have not been independently verified — say that plainly, because the gap between a vendor’s headline and a user’s experience is exactly where disappointment lives.

What independent evidence exists so far is thinner but encouraging. One early timing, published in launch-week coverage, measured a medium-quality image at 4.84 seconds on version 2.6 against 7.08 seconds on version 2.5 — a real improvement, but a single measurement from one user, not a benchmark. Treat it as a data point, not a promise. Speed also depends on the model and the settings you choose: Flare at low quality is the fastest combination OpenAI offers, while Sunburst at high quality with high complexity will always be slower, because it is doing more work. Higher complexity, by OpenAI’s own warning, increases latency.

For most ChatGPT users, the speed gain will feel modest rather than dramatic. A few seconds saved per image matters when you generate fifty drafts in an afternoon; it barely registers when you generate one. The users who will feel it most are the ones Flare was built for: developers running high-volume generation, and creators iterating fast on a visual idea. For them, a 30% to 50% cut in generation time, repeated hundreds of times a day, is a genuine workflow improvement — and it is the main reason to choose Flare over Sunburst when quality is not the deciding factor.

GPT Image 2.6 vs GPT Image 2.5: What Actually Changed

Feature GPT Image 2.5 GPT Image 2.6
API models One model (gpt-image-2.5) Two models: Flare (speed) and Sunburst (detail)
Speed Baseline About 50% faster (OpenAI); Flare as fast as 6s at low quality
GIF creation Not available Yes — ChatGPT and API; 2–64 frames, adjustable speed, editable frames
Saved references Re-upload each time Save once, reuse with @name
Image markup ChatGPT only ChatGPT and API
SVG export Not available API only, static vector, one image at a time
Complexity setting Not available Low / medium / high; default high
Token rates $5 / $8 / $30 per 1M tokens Identical rates
Cost per low-quality image ~$0.006 ~$0.025 (up to 4.2× more output tokens)
Cost per high-quality image ~$0.053 ~$0.053 (unchanged)
Batch API Supported at half price Not supported
Availability ChatGPT, Codex, API ChatGPT, Codex, API — every ChatGPT tier

The table makes the trade-off clear. GPT Image 2.6 is a better product in almost every dimension that a user can see: faster, more consistent, animated, and more precise. It is also a more expensive product at the quality settings where most bulk work happens, and it takes away the batch discount that made bulk work cheap. Whether the upgrade is worth it depends entirely on which side of that trade-off your work sits on. If you make a handful of careful images a day, 2.6 is strictly better. If you generate thousands of cheap drafts a day, you should re-run your budget before migrating.

How to Use GPT Image 2.6 in ChatGPT

Using GPT Image 2.6 in ChatGPT needs no setup: the new models are the image models ChatGPT uses now, on every plan from Free to Pro, on web, desktop and mobile. You do not pick Flare or Sunburst in the chat window — that choice exists only in the API. What you do in ChatGPT is use the new abilities.

  1. Generate an image as usual. Describe the image you want in plain language. ChatGPT now uses Images 2.6, so generations are faster and edits hold together better than they did on version 2.5.
  2. Make a GIF. Ask for an animation — “turn this into a looping GIF of the mascot waving” — or upload a photo or a sketch and ask ChatGPT to animate it. Adjust the playback speed, and edit individual frames if a frame looks wrong. If the GIF option does not appear yet, it is still rolling out to your account.
  3. Save a reference. Upload a photo of yourself, your pet or a character, save it as a reference and give it a name. In any later conversation, type @ followed by that name to reuse it instead of uploading the photo again.
  4. Edit precisely. When something needs fixing, describe the exact change — or point at the area in the image — rather than re-describing the whole picture. The model is built to hold the rest of the image steady now.
  5. Ask for the right format. If you need a picture for a post, say so; if you need a transparent background or a specific aspect ratio, state it in the prompt. ChatGPT cannot return an SVG, so designers who need vector files should use the API instead.

A practical tip from the launch coverage: for edits, ask for one change at a time. The model’s edit consistency is better than version 2.5, but a prompt that asks for five changes at once still gives it five chances to change something you liked. One change, check the result, then the next change — that is slower per step and faster overall, because you throw away fewer images.

How Developers Use the GPT Image 2.6 API

Developers reach GPT Image 2.6 through the Images API, and the first choice is the model identifier: gpt-image-2.6-flare for fast iteration, gpt-image-2.6-sunburst for detailed work and consistent edits. Both models support the image generation and image edit endpoints, and both accept text and image input with image output.

The new controls are set per request. Complexity takes low, medium or high and defaults to high — set it explicitly rather than accepting the default, because high complexity costs latency and can introduce changes you did not ask for. For output format, output_format accepts the usual raster formats plus gif and svg: GIFs take a frame count (2 to 64, default 9) and a frame duration in milliseconds (20 to 2,000, default 200), while SVG requires exactly one image per request. Markup is passed as a region selection on the edit endpoint, letting your app turn a user’s tap on part of an image into a precise edit instruction.

A developer writing code on a laptop, integrating the GPT Image 2.6 API into an application

Three integration notes from the launch documentation are worth building into your plan. First, Batch is not supported for either 2.6 model, so a pipeline that relied on half-price batch generation on 2.5 needs a new cost model — or should stay on 2.5 for bulk work. Second, the model pages mark the Responses API as unsupported for 2.6, so if your integration used the Responses tool with 2.5, verify the migration path before you switch. Third, GIF output in the API was behind an undocumented feature flag at launch, so do not build a product feature on API GIFs until OpenAI documents and enables it generally.

Pricing in the API is per token, not per image: $5 per million text-input tokens, $8 per million image-input tokens and $30 per million image-output tokens, with cached input at $1.25 and $2 respectively. Budget on the measured token counts in the pricing table above, not on the version 2.5 counts your existing code assumes — a low-quality image now uses roughly 823 output tokens instead of roughly 200, and your cost monitoring should reflect that from day one.

Flare or Sunburst: Which GPT Image 2.6 Model Should You Pick?

The choice is simpler than the two-model launch makes it look, because the models cost the same. You are choosing between waiting and quality, and the right answer depends on where each image sits in your workflow.

  • Choose Flare when you are drafting, brainstorming or generating in bulk. Product catalogues, social media variations, poster layouts, first-pass storyboards and any job where you will make ten images and keep one. Set quality to low or medium and complexity to low, and let the speed do the work.
  • Choose Sunburst when the image is the final asset. The hero image on a landing page, a brand visual that must survive five rounds of edits, a detailed illustration with text, or any crowded composition where many elements must hold together. Set quality to medium or high and complexity to medium or high, and let the detail do the work.
  • Choose Sunburst for edits, even if you drafted in Flare. This is the workflow the two-model split is built for: draft fast in Flare, then hand the keeper to Sunburst for the careful final passes. The models share the same rate card, so the split costs you nothing extra per token — it only costs you the discipline to switch.
  • Choose by quality setting, not by habit. If you always generate at high quality, the models’ cost is identical and Sunburst’s consistency is free. Flare’s advantage only appears at low and medium quality, where speed is the point.

Limitations and Honest Caveats

No launch-day article about GPT Image 2.6 is complete without the parts OpenAI did not emphasise. Here they are, plainly.

The speed claim is OpenAI’s, not an independent benchmark. “About 50% faster” comes from the company, and the only independent timing published in the first days was a single user’s measurement. Your speed will vary with model, quality, complexity and load. Do not promise a client a six-second image until you have timed it on your own account.

Low-quality images cost more than they did. This is the most important caveat in the article, and it bears repeating: token rates are unchanged, but output token usage at low quality rose up to 4.2 times, so a low-quality image went from about $0.006 to about $0.025. At high quality nothing changed. Budget accordingly.

GIFs are still rolling out, and the API flag is off. In ChatGPT, GIF creation was still reaching users gradually at launch. In the API, GIF output sat behind an undocumented feature flag. And ChatGPT cannot edit an animated GIF you upload. If GIFs are the reason you are switching, confirm the feature is live on your account first.

SVG is API-only and one-directional. You cannot get an SVG from ChatGPT, you cannot ask for more than one at a time, and you cannot feed an SVG back into the edit endpoint. The vector you get is static, with no scripts or embedded raster images. It is a starting point for a designer, not a finished design file.

Batch pricing is gone for now. Version 2.5 had a half-price Batch rate. Neither 2.6 model has one. Bulk pipelines should stay on 2.5 — or be re-budgeted — until that changes.

Likeness and watermark questions are unanswered. Saved references and GIFs make it easy to animate a real person’s photo. OpenAI’s announcement does not explain what protects real people’s likenesses, or whether GIFs of people carry a watermark. Use reference images only of yourself, people who have agreed, or characters you own.

Leaderboard ranks are early signals. Sunburst’s first-place rank on DesignArena’s Image Arena and on Artificial Analysis arrived on launch day, when few independent testers had run the model. Leaderboards are useful as a snapshot, not as a verdict, and the editing ranks in particular were close.

GPT Image 2.6 vs the Competition

GPT Image 2.6 does not launch into an empty market, and its rivals frame what it is and is not. Canva’s Craft, announced in late October 2026, is the design-first alternative: a model that plans a layout before it draws, with legible text in 55 languages and a public API, aimed at the same designers who will care about SVG export. The open-weight route is represented by models like FLUX 3 Image Klein, which trades OpenAI’s convenience for weights you can run yourself — the comparison is covered in our FLUX 3 Image Klein guide. And for the video-adjacent work that GIFs only partly cover, the full video-generator landscape is in our best AI video generator roundup, which remains the right starting point if what you actually need is a moving picture with a camera, not a loop.

What distinguishes GPT Image 2.6 from those rivals is not any single feature. It is the combination, inside the tool hundreds of millions of people already use: animation, character memory, precise edits and vector output, all in ChatGPT, with a two-model API behind it. Canva Craft is arguably the stronger design tool, and open-weight models are arguably the cheaper long-term bet for self-hosters. But neither of them is already sitting inside the chat window where the work happens. For most US creators and small teams, that distribution advantage — plus the 4-billion-images-a-week scale that keeps improving the model — is the real reason GPT Image 2.6 will be the default image tool they reach for, even where a specialist rival is technically better at one job.

GPT Image 2.6 FAQ

What is GPT Image 2.6?

GPT Image 2.6 is OpenAI’s November 2026 image model, released on 4 November in ChatGPT, Codex and the API. It adds GIF creation, saved reference images, faster generation, image markup in the API, SVG export and a complexity setting, split across two API models — Flare for speed and Sunburst for detail.

What is the difference between Flare and Sunburst?

Flare (gpt-image-2.6-flare) is built for fast iteration and can generate an image in as little as six seconds at low quality, according to OpenAI. Sunburst (gpt-image-2.6-sunburst) is built for detailed work and consistent edits at medium or high quality. Both cost the same per token, so the choice is seconds versus detail, not price.

How much does GPT Image 2.6 cost?

In ChatGPT, image generation is included in your subscription on every plan, with no per-image price. In the API, both models cost $5 per million text-input tokens, $8 per million image-input tokens and $30 per million image-output tokens. Because version 2.6 uses more output tokens per image, a low-quality 1024×1024 image costs about $0.025 (up from about $0.006 on version 2.5), a medium one about $0.043 and a high-quality one about $0.053.

Can GPT Image 2.6 make GIFs?

Yes. In ChatGPT you can turn a prompt or an image into a looping animated GIF, adjust its speed and edit individual frames. In the API, GIFs run from 2 to 64 frames with adjustable frame duration. GIF creation in ChatGPT was still rolling out at launch, and ChatGPT cannot edit an animated GIF you upload.

Can GPT Image 2.6 export SVG files?

Yes, but only through the API. Set the output format to “svg” and request one image at a time. The SVG is a static vector file with no scripts, external resources or embedded raster images, and the edit endpoint cannot return or accept an SVG.

Is GPT Image 2.6 available on the free plan?

Yes. OpenAI says Images 2.6 is available in ChatGPT on every subscription tier, from Free to Pro, on web, desktop and mobile, and in Codex. GIF creation was still rolling out at launch, so it may not have reached every account on day one.

Should I switch from GPT Image 2.5 to 2.6?

For most users, yes — 2.6 is faster, more consistent and more capable. For API users running bulk low-quality generation, check the new token counts first: low-quality images cost about four times more than on 2.5, and the Batch half-price rate is not available for 2.6. If your pipeline depends on batch pricing, stay on 2.5 until that changes.

The Bottom Line

GPT Image 2.6 is the most practical image update OpenAI has shipped in a while, and the reason is not the version number. GIFs give ChatGPT users a new medium. Saved references fix the character-consistency problem that made multi-image work frustrating. Markup and the complexity setting give editors the precision they asked for. SVG export gives designers a file they can actually use. And Flare and Sunburst give developers a real choice between speed and care, at the same price per token.

The catch is the price — not the rate card, which did not change, but the token usage underneath it. Low-quality images now cost about four times what they did, and the batch discount is gone. That does not make GPT Image 2.6 a bad deal; at high quality it costs exactly what version 2.5 cost, and it does more. But it does make the upgrade a decision rather than a reflex, especially for anyone generating images by the thousand.

If you use ChatGPT, the decision is already made for you: GPT Image 2.6 is the image model now, on every plan, and the GIFs and saved references are worth trying today. If you build on the API, draft in Flare, finish in Sunburst, set complexity explicitly, and re-run your cost maths on the new token counts before you migrate a bulk pipeline. Do that, and GPT Image 2.6 is what it promises to be — the fastest, most consistent image model OpenAI has shipped, with the one honest surprise being the bill, not the pictures.

Leave a Comment

Your email address will not be published. Required fields are marked *