ChatGPT Makes GIFs Now: OpenAI Ships Images 2.6 with Animation, Two Models, and 50% Faster Generation

On November 4, 2026, OpenAI released Images 2.6. The headline is plain enough: pictures can move now. Users can generate animated GIFs directly in ChatGPT, Codex, and the API, then edit them frame by frame and adjust playback speed. At the same time, the API gains two new models — GPT-Image-2.6 Flare, tuned for speed, and GPT-Image-2.6 Sunburst, tuned for detail — and OpenAI says image generation is 50% faster than the previous generation. This is not a new foundation model. It is a practical capability upgrade: animation, steadier multi-turn editing, and finer API controls. For creators and indie developers who make images every day, it fixes several of the most annoying small problems.

The numbers first: 4 billion images a week

OpenAI opens its announcement with a scale figure: people now create 4 billion images a week across ChatGPT, Codex, and the API — more than half a billion a day on average. In September that figure was 3 billion a week, so it grew by about a third in two months. Image generation stopped being a novelty a while ago; it is now one of the most-used features in ChatGPT. That context matters, because every change in Images 2.6 is aimed at people who already use it daily, not at demo audiences. GIFs, reusable reference images, and SVG export are not showpiece features. They fill gaps in a production workflow.

The rollout scope is also worth noting: ChatGPT and Codex on all tiers, on web, desktop, and mobile, globally, with both new models in the API on the same day. Regular users and developers got the new capabilities at once, with no staggered teasing. That said, following OpenAI's usual pattern, a user-facing feature like GIF creation may still arrive on different accounts at different times, so treat "available today" as the announcement's wording and check what your own account actually shows.

What actually shipped

Broken down, Images 2.6 is five things in one package. First, GIF animation: looping animations from a text prompt, a sketch, or an existing image, with adjustable playback speed and frame-by-frame inspection and editing. Second, two new API models, Flare and Sunburst, with a clear division of labor — one for speed, one for detail and consistency. Third, a new complexity parameter that lets developers control how much planning the model does before it draws. Fourth, image markup editing in the API, so you can point at exactly where a change belongs. Fifth, SVG export, so generated designs can go straight into websites, apps, and print.

There is also an easily missed addition: reusable reference images. Once a character, pet, or person is set, it can be reused as a reference so later generations keep the same look. That matters a lot for anyone producing a series — previously, every image was a fresh roll of the dice for the same character, and now there is at least an official mechanism for keeping it stable.

GIFs: from still images to looping motion

GIFs are the most visible feature of this release, and also the easiest to misunderstand. OpenAI's positioning is clear: this is not video generation. A GIF is a short looping animation, suited to animating a brand mascot, making reaction images and memes for social media, or building animated storyboards to explore character movement before video production. The examples in the announcement are concrete: marketers can animate a mascot for a campaign, creators can make reaction GIFs, and video teams can test character motion before committing to full production.

The two example prompts in the announcement also reveal how the GIFs are made: the model first generates a sprite sheet — six sequential frames arranged in a precise 3-column by 2-row grid — which the product then plays as a loop. The first example is an office worker watching a stock-market line plunge: hopeful expression, noticing the fall, eyes widening, face crumpling, big cartoon tears, then quietly wiping them away while still staring at the chart. The second is the mirror image, a soaring chart and a celebration, grabbing banknotes and tossing them like confetti. Both require the sixth frame to transition smoothly back to the first — a seamless loop.

That method is worth pausing on. It suggests Images 2.6 GIFs are closer to "moving picture books" than to true frame-by-frame video: the range of motion and camera movement is limited by the six-frame single-image structure. The upside is control, low cost, and fast output; the price is that you should not expect fluid camera moves or complex action. For indie developers, the genuinely useful line in the announcement is a quieter one: even a simple sketch can become an expressive animated character. The announcement shows an owl drawn in pen on lined paper, making the point that you do not need to be able to draw to start. Anyone making games, sticker packs, or a product mascot should immediately see a use for that.

OpenAI names one early user: Town, a startup building an AI work assistant, which used Images 2.6 in the API to explore how animated characters could add personality to everyday interactions at work. That is a signal: the first real users of GIFs are not ad agencies, but software teams treating "personality" as a product feature.

How the two models divide the work: Flare for speed, Sunburst for detail

The API-side changes deserve a closer read than the GIFs in the ChatGPT interface. Images 2.6 splits the model line in two: GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst. According to OpenAI, Flare is the fast one — as little as about six seconds per image when quality is set to low — and it also improves layout quality for posters, product imagery, and other professional visuals. Sunburst is the detailed one: at medium or high quality it produces richer detail and holds the subject steadier across a chain of edits.

This continues the direction set by Images 2.5 in September, when Flare and Sunburst first appeared as a pair. Version 2.6 is an upgrade inside that same framework. The practical question for developers is simple: does this image need to be fast or steady? Try directions and generate candidates in bulk with Flare; finalize, build a series, or edit the same image repeatedly with Sunburst. Both models are billed the same way, per token, so the cost of choosing is low — the expensive mistake is picking the wrong one and redoing the work.

As for "50% faster," that is OpenAI's overall figure against the previous generation, with no detail on which quality setting or image size it was measured at. Third-party analysis has pointed out that OpenAI's own published Flare benchmark shows elapsed time falling from 7.08 seconds to 4.84 seconds — roughly a 32% improvement, not 50%. The two numbers may measure different configurations, but at minimum they show that "50% faster" should not be read as "every image takes half the time." Your actual speedup depends on your settings.

Four changes developers should actually look at

For teams calling the API, GIFs are the garnish. The four changes below are the reason to upgrade.

1. The complexity parameter: decide how hard to think before drawing

The new complexity parameter has three settings — low, medium, and high — and controls how much planning the model does to meet a request. It is not a quality dial: quality controls visual detail and fidelity, while complexity controls how much goes into the scene. OpenAI's advice is direct: use low complexity for most editing tasks, and try medium or high for scenes with many elements or constraints. In developer terms, this is a cost-versus-effect switch — do not make the model think hard about a simple edit, and give it time for a complex layout. For pipelines that process images in bulk, this parameter may save more money than the model upgrade itself, because it lets you pay for the difficulty of the task rather than running everything at the top spec.

2. Markup-guided editing in the API: circle what to change

Image markup was previously only available in ChatGPT. Now it comes to the API: alongside the source image, you send a marked-up overlay showing exactly where the change belongs. That addresses the most common frustration in editing calls — "I said change the logo in the top left, and it changed the text in the bottom right." For teams automating e-commerce images, product shots, or posters, markup editing replaces the unreliable "describe the location in words" approach with "point at it directly." Whether it truly changes only the marked spot and leaves everything else alone will need real-world testing, but the direction is right.

3. SVG export: generated images that enter the design pipeline

The API can now export SVG vector output. For logos, icon sets, and anything headed for print or a design pipeline, this matters: generated images used to come back as raster files that went soft when scaled, forcing designers to re-trace them. Vector output can now be used directly on websites, apps, and printed material at any size without losing sharpness. For indie developers and small teams, this means AI-generated images can finally be handed over as usable design assets, not just mock-ups for showing what something might look like.

4. Reusable reference images: stop re-rolling the same character

The hardest part of multi-turn editing is consistency: edit the same character five times and it comes back looking different five times. Images 2.6 attacks that in two ways: the Sunburst model itself holds the subject steadier across edits, and the reusable reference image mechanism lets characters, pets, people, and brand elements be fixed and reused. For anyone producing a series, brand visuals, or game assets, this is the step from "one image at a time" to "a coherent set." The announcement does not state limits on how many reference images can be stored, similarity boundaries, or copyright filtering details — those answers will have to come from the API documentation and hands-on testing.

Which numbers are vendor claims, and which are verified

This is where the accounting has to be honest, because almost every striking number in this launch comes from OpenAI itself.

"50% faster" is a vendor claim with no independent verification yet. As of November 5, Artificial Analysis's text-to-image and image-editing leaderboards did not list GPT-Image-2.6 Flare or Sunburst; the top of the board was still the previous generation, GPT Image 2.5 Sunburst and Flare. In other words, the new models have no public benchmark score yet, and any judgment that they are "better" can only rest on OpenAI's word for now. "As little as six seconds" for Flare is likewise conditional — it applies to the low-quality setting, and the announcement says nothing about how long medium- or high-quality generations take.

"4 billion images a week, more than 500 million a day" is also OpenAI's own platform figure, impossible to verify from outside. What it signals is a trend rather than a precise count: 3 billion a week in September to 4 billion in November is about a third of growth in two months, which suggests demand for image generation is still climbing fast. If the figure is accurate, it means every small improvement in Images 2.6 is multiplied across hundreds of millions of calls, and API stability and cost control will become the real test sooner than the features themselves.

On pricing, the announcement gives no new price list, only a link to the pricing page. Under the API documentation's existing rules, the two new models continue to bill per token rather than per image, at the same rates as Images 2.5. So this upgrade neither raises nor lowers prices — but the billing details for GIFs through the API (whether they are charged per frame or per duration) are not spelled out either, so read the pricing page before planning around them.

What it means for people building products

For VibeFix readers — indie developers, side-project builders, small teams — the value of Images 2.6 is not "another model." It is three concrete scenarios.

First, the last mile of product visuals gets shorter. AI-generated landing page art, social assets, and product screenshots used to suffer from two problems: changing one thing moved everything else, and the style would not stay consistent. Markup editing plus reusable reference images at least give you a mechanism for "finalize, then fine-tune," and SVG export means logos and icons can be used as-is. Builders can spend more time on what an image should say and less time re-rolling.

Second, the barrier to motion has dropped to the floor. A GIF needs no video model and no editing software — one sentence plus one image can make a mascot blink or a product screenshot come alive. On a homepage, a README, or a social feed, a looping animation often holds attention better than a still image. Notably, OpenAI's own examples include "turning a product shot into an eye-catching loop," which reads like a direct message to indie developers.

Third, the cost structure is clearer. The Flare-versus-Sunburst split, plus the complexity parameter, amounts to a pay-for-what-you-need menu: cheap and fast combinations for trying directions, high spec only for finals. After this release, running everything on the strongest model looks wasteful. Knowing how to use that menu may decide whether your image bill goes up or down.

Questions the announcement leaves open

What the announcement does not say matters as much as what it does. First, the GIF limits are not public: maximum duration, frame count, and resolution caps are not stated, and API documentation is not fully consistent on them — developers should verify before integrating. Second, there is no promised rollout schedule for GIFs in ChatGPT; arrival times may differ by days or weeks across accounts and plans. Third, the boundaries of reusable reference images are unclear: how many can be stored, what the review rules for people's likenesses are, and how commercial rights work are all questions a shipping team needs answered first. Fourth, the announcement gives no comparable data on how much better Sunburst's multi-edit consistency is than 2.5's — that will have to wait for hands-on testing.

There is also a longer-term question: are GIFs just a stand-in for video generation? OpenAI deliberately keeps GIFs separate from video in this announcement, emphasizing short loops that are light and shareable. But once users get used to images that move, the next question is inevitable: can they move longer and more smoothly? Images 2.6 answers today's question. The video side of that story lives elsewhere.

Three things worth trying right away

If you want to try it now, three experiments are cheap and informative. First, take an existing product screenshot or mascot image and generate a GIF, then see whether the loop is good enough for your homepage or social posts. Second, run the same repeatedly edited scene through Flare and Sunburst and compare speed, consistency, and the bill before choosing a production line. Third, pick an image that needs a local change and try markup editing, to test whether "circle it and change it" really beats describing the spot in words. None of these requires changing your architecture, all three can be done in half a day, and together they answer whether Images 2.6 belongs in your workflow.

Verdict: a capability upgrade, not a generational leap

Images 2.6 is not a new foundation model, and OpenAI does not present it as one. It builds on Images 2.0 (April 2026) and the Flare/Sunburst split of Images 2.5 (September 2026), filling the gaps in the production workflow in one go: animation, markup editing, SVG export, and complexity control. It should be judged by productivity standards, not by how spectacular it looks.

For regular users, GIFs are the most direct win: memes, mascots, and storyboards with a barrier so low that a sketch is enough to start. For developers, the real value is the four API-side controls — the complexity parameter, markup editing, SVG export, and reusable reference images — plus the clear Flare/Sunburst division of labor. Together, they move image generation closer to production and further from pulling a slot machine.

But stay clear-eyed: every key number — 50% faster, six seconds, 4 billion images a week — is self-reported by OpenAI, and the new models have no public leaderboard scores yet. The sensible approach is to measure Flare on your real tasks for speed and cost, measure Sunburst on your repeated edits for consistency, and only then decide whether to move your production pipeline over. The features are live; the verification is up to the people using them.