How to Make GIFs in ChatGPT With Images 2.6

On 4 November 2026, OpenAI gave ChatGPT a new trick: the ability to turn a sentence or a photograph into a short animated GIF. The feature arrives with Images 2.6, the latest version of the image system behind ChatGPT and Codex, and it works alongside two new models for developers. This guide explains how to make GIFs in ChatGPT, what controls you actually get, how reference images keep a character consistent, and where the limits are.

What Images 2.6 is

Images 2.6 is OpenAI’s image-generation release for November 2026, succeeding Images 2.5 from September. It is not a chatbot model or a video generator. It is the system that draws pictures inside ChatGPT and Codex, now extended from still images into short, looping animation. OpenAI says ChatGPT and its API together now produce more than 4 billion images a week, up from 3 billion in September — which explains why even a seemingly small addition, like GIFs, reaches an enormous audience immediately.

The release has three parts: GIF creation inside ChatGPT and Codex, reference images for reusing subjects, and two new API models for developers. Each is covered below.

How to make a GIF, step by step

Making a GIF takes the same kind of request you would already use for a still image, plus a motion description:

1. Start with a prompt or a photo. Either describe the scene and the movement in plain language (“a cat dancing in a kitchen, looping”) or upload a photo or sketch you want animated.
2. Describe the motion, not just the subject. The model needs to know what moves and how: waving, bouncing, spinning, reacting. Vague motion gives vague animation; a specific action gives a cleaner loop.
3. Let it loop. The output is a short animation designed to repeat seamlessly. Check the seam — the point where the end jumps back to the start. A good loop hides it; a bad one visibly stutters.
4. Adjust the speed. Reports of the rollout put playback control between 0.5x and 2x. Slower speeds suit gentle motion like a waving mascot; faster speeds suit punchy reactions.
5. Edit individual frames. If one frame breaks the illusion — a hand that warps, a face that shifts — you can revise or replace that frame without regenerating the whole GIF.
6. Share it on repeat. The finished GIF is a standard looping animation, ready for a social post, a chat reply, or a product page.

If GIF creation is not visible in your account yet, that is expected: OpenAI says the rollout is gradual, even though it covers every tier including free. There is no separate paywall for GIFs, but normal account usage caps still apply, so heavy use will still hit the usual limits.

What makes a good GIF prompt

Short animation rewards simplicity. A single subject, a single action, and a static background produce the cleanest loops — which is why OpenAI’s own demonstrations feature a dancing cat and a cartoon office worker reacting to a stock chart, both holding character, camera and background steady across frames. Busy scenes with several moving subjects, camera moves, and changing backgrounds are where loops break down: the model has to keep too many things consistent at once.

A practical rule: write the prompt as subject + action + setting, in that order, and keep the action to one verb phrase. “A red mascot waving, plain background” will loop more cleanly than “a mascot running through a market while waving at stalls and dodging carts”. If the first result stutters, simplify the motion before blaming the tool.

Reference images: keeping a character the same

The second major addition solves a long-standing annoyance with AI images: every new picture of the same character used to come out slightly different. With reference images, you save pictures of yourself, a pet, a character, or a product once, and then reuse them across prompts without uploading them again.

This matters most for repeated content. A brand mascot that must look identical in every post, a comic character across a series, or a product shot that must match the real item — all of these become practical once the model has a saved reference to anchor to. Combined with GIFs, the workflow is: save the reference once, then generate a new looped animation of the same character whenever you need one, with the speed and frame controls described above.

One caution applies. A saved reference makes it easier to produce convincing images of real people and pets, so use your own images or ones you have permission to use, and expect the usual platform rules about impersonation to apply.

Flare vs Sunburst: the two API models

Developers get two new models through OpenAI’s API, continuing the two-model approach from Images 2.5:

• GPT-Image-2.6 Flare — tuned for speed and high-volume, everyday creation: interactive interfaces, bulk requests, and anything where latency matters more than pixel perfection. OpenAI’s published Flare benchmark shows generation falling from 7.08 seconds under Images 2.5 to 4.84 seconds under 2.6.
• GPT-Image-2.6 Sunburst — tuned for quality and edit consistency: product imagery, character studies, and assets that go through several revisions, where the tenth edit must still match the first. Sunburst also adds SVG export, useful for logos and graphics that need to scale cleanly.

OpenAI’s headline claim is that the new API models generate images 50% faster than their predecessors. The Flare benchmark above implies a smaller time saving on that particular measurement, so treat the headline as a best-case figure rather than a uniform speedup. Both models also gain ChatGPT’s image-markup tools in the API, letting a team circle an exact region of an image when asking for a change, instead of describing the region in words. Pricing for Flare and Sunburst was not included in the announcement, so developers should check the current API pricing page before estimating costs.

GIFs are not video — and that is the point

It is tempting to compare GIF creation with video generators, but the two do different jobs. A video generator aims for a scene: camera movement, narrative, seconds or minutes of changing action. A GIF is a few seconds of repeated motion with no story arc. OpenAI frames Images 2.6 the same way, as short looping motion rather than long-form generation.

The distinction matters for expectations. If you want a product demo, an explainer, or a short film, a GIF tool will disappoint. If you want a reaction image, a mascot wave, a loading animation, or a storyboard sketch that moves just enough to show the idea, a GIF is the lighter, faster, more shareable format — and it plays everywhere, including in chats and emails where video often does not autoplay.

Availability and limits

• Where: ChatGPT and Codex, on the web, desktop and mobile.
• Who: all subscription tiers, including the free plan, with GIF creation rolling out gradually.
• Caps: no separate GIF paywall, but existing account-level usage caps apply.
• Editing: speed control and per-frame edits are built into the chat workflow.
• API: Flare and Sunburst available to developers; pricing not disclosed at announcement.
• Output: short loops only — not a replacement for video generation.

Common problems and how to fix them

Short animation fails in predictable ways, and most of them have simple fixes:

• The loop visibly jumps at the seam. Simplify the motion or the background. A plain background and a motion that returns to its starting pose — a wave that goes up and comes back down — hide the seam far better than a walk cycle or a camera pan.
• The character drifts between frames. Save a reference image of the character first, then generate the GIF from that reference instead of describing the character fresh each time. If drift persists, reduce the number of moving elements in the scene.
• Text in the GIF comes out garbled. Keep on-image text to a few short words, or leave text out of the animation entirely and add it afterwards in an editor. Small text is the first thing short animation mangles.
• The motion feels too fast or too frantic. Drop the speed toward 0.5x before regenerating. A slower loop reads as deliberate; a fast loop of a complex motion reads as a glitch.
• One frame breaks the whole clip. Use the per-frame edit rather than regenerating. Replacing a single bad frame preserves the frames that already work, while a full regeneration rolls the dice on all of them again.

As a general rule, fix the prompt before blaming the model: one subject, one action, one static background. Complexity is the enemy of a clean loop.

What to watch next

Three things will decide how important Images 2.6 turns out to be. First, how the free tier actually behaves once the gradual rollout completes — free users are the largest audience, and their experience, including caps, will define the feature’s reach. Second, whether reference images stay consistent over long series of edits, which is the promise Sunburst in particular makes. And third, what creators do with the format: the GIF spent two decades as the internet’s reaction language, and putting its production inside the world’s most-used chatbot may revive it as a branding and communication tool rather than just a joke format.

For the day-by-day news version of this story, see our report: ChatGPT Can Now Make GIFs: OpenAI’s Images 2.6 Turns Prompts and Photos Into Animation. OpenAI’s own announcement is the primary source for this guide: Images, upgraded — introducing ChatGPT Images 2.6.