OpenAI shipped Images 2.6 to ChatGPT, Codex and the API on November 4. The feature most people will try first is a GIF. The part that changes daily work is speed, plus an API that now splits into two models.
What ChatGPT Images 2.6 adds on day one
The update is live globally on every subscription tier, and it runs on desktop, mobile and the web. OpenAI leads its announcement with animated GIF creation, which tells you where the company expects attention to land. The same model also shows up in Codex and in the API, so a feature you first meet in a chat window is available to anyone building on top of it.
OpenAI published a system card for Images 2.6 as well. That document is where to look if you want to know how the model was evaluated, what limits the company set for itself, and where those lines fall. It won't tell you whether the GIFs are good, but it sets expectations for what the model won't do.
The ChatGPT GIF maker, minus the demo gloss
OpenAI's examples cover three jobs: animating a mascot, turning a product shot into a loop, and building an animated storyboard. All three are short repeating visuals, which is the shape a GIF wants. A loop has to return to its starting point, and that constraint should shape what you ask for.
For a normal user, the workflow change is smaller than the launch post suggests. You stay in the chat. You describe the motion in one sentence. You ask for revisions in plain language. There's no timeline, no keyframes, and no export into a motion editor, which is either the feature's best quality or its main limitation, depending on your job.
The motion itself gets inferred. You don't set a frame count or a duration, and you don't drag a playhead. You describe an action in words, judge the result, then describe the correction. For small loops that beats a timeline. For anything with real timing cues, it's guesswork with a preview.
The 50% speed claim OpenAI attached to Images 2.6
OpenAI says Images 2.6 generates images 50% faster than the previous version, and that people create 4 billion images per week across ChatGPT, Codex and the API. Both numbers come from the company. Neither has been checked by an outside lab, and no independent benchmark of the speed claim exists yet.
Keep the two figures apart. One compares two versions of the same OpenAI image model. The other measures how heavily the products get used.
Two very different claims.
Does a 50% cut change what people ask for? Probably, and in a boring way. Waiting is the tax on iteration, so when the tax drops, people run more attempts before they settle. A tool that felt slow starts to feel normal.
GPT-Image-2.6 Flare and Sunburst split the API in two
The API now offers two variants, and the difference is a straight trade between speed and finish.
Variant | Tuned for | Where it fits |
|---|---|---|
GPT-Image-2.6 Flare | Speed, with runs as short as 6 seconds at low quality, and better layout for posters and product imagery | The stage where you're still deciding what the image should be |
GPT-Image-2.6 Sunburst | Higher quality and stronger consistency across edits at medium and high quality | The stage where the layout is settled and each revision has to hold |
Flare is the one for exploring. Sunburst is the one for finishing, and the names hint at the split: a burst of options against the frame you keep. OpenAI's own description backs it, with Flare measured in seconds at low quality and Sunburst tuned for medium and high quality where edits have to stay consistent.
What else changed in the GPT-Image-2.6 API
Three additions sit below the headline. A complexity parameter now sets how much planning the model does before it draws, so you can spend latency on care or skip that step. Image markup that used to live only in ChatGPT now reaches the API. SVG export is supported too, which matters to anyone who needs a file that scales without pixelating.
For developers, the markup carry-over may be the useful one. A prompt you tuned in the chat window can move into code with fewer rewrites.
The complexity parameter matters more if you run batches than if you make one image at a time. Planning costs seconds, and seconds cost money once you're running thousands of images, so the setting turns into a budget decision instead of a quality one.
Where the ChatGPT GIF maker still gets awkward
Motion is the hard part, not the first frame. A bouncing mascot is easy. A character that has to keep the same face across eight frames isn't, and frame-to-frame consistency is where most attempts fall down.
People who want precise motion often describe a sprite sheet, meaning several frames arranged in a grid that they assemble outside the chat. It works. It's also tedious, and it hands the timing decisions back to you.
One limit deserves a plain sentence: this isn't video, and Images 2.6 makes animated GIFs, which are short loops by design. There's no audio, no scene changes, and no two-minute explainer waiting at the end of it.
Who will notice Images 2.6 first
Social teams have the clearest win, because a looping product shot costs less to produce than a filmed one. Teachers and explainer writers get a second: a short animation of a process often beats a still diagram for the same idea.
Agencies get a third. A client who wants one product in six colours, moving, is now a short job, not a studio booking.
Designers will care more about the API split than about the chat feature. Flare cuts the cost of exploring options, while Sunburst keeps a settled layout from drifting across a long series of revisions.
What to watch after the Images 2.6 rollout
Two things. Watch for an independent speed test, since the 50% figure is still a vendor claim. Then watch how GIF output holds when a project needs ten variations of one character across a campaign, which is the point where consistency stops being a demo trick.
Images 2.6 is live now across ChatGPT on desktop, mobile and web. Open a chat, ask for a loop of an image you already have, then ask for one change. That's the fastest way to see whether the feature earns a place in your week.
