How to Use ChatGPT Images 2.6: Make GIFs, Edit Faster, Save References

ChatGPT Images 2.6, launched on 4 November 2026, turns a typed prompt or a still photo into a looping GIF in a single step. This guide on how to use ChatGPT Images 2.6 walks through making that first GIF, changing its speed, fixing a single frame, and sharing the result — then covers faster still images, saved photo references, and the two API models, Flare and Sunburst. One warning comes first: GIF creation is still rolling out, so some accounts will get a still image instead of a GIF. That is normal for now, and there is a workaround further down.

Read this first: GIF creation is still rolling out — if you get a still image instead of a GIF, your account doesn’t have it yet; use the sprite-sheet method below.

How to use ChatGPT Images 2.6: make your first GIF

OpenAI’s own demo prompt is a good place to start: “can you make a gif of a cat dancing?” Type a prompt like that into ChatGPT, or describe any short scene you want to see move, and the model builds the loop for you. There is no separate GIF mode to switch on and no timeline to draw — one prompt produces the finished loop.

1. Open a fresh chat in ChatGPT on desktop, mobile, or web. OpenAI says Images 2.6 is available globally across all three, and across all subscription tiers.
2. Describe the motion. “A cat dancing” works better than “a cat”, because the model needs to know what should move. Name the subject, the action, and the setting in one short sentence.
3. Wait for the loop to play. The GIF arrives as a looping animation you can watch straight in the chat.
4. Adjust the speed. Playback speed is adjustable, so slow a fast dance down or speed a slow wave up until the loop feels right.
5. Save or share it. The GIF can be saved to your device or shared from the chat, the same way you would share any image ChatGPT makes.

A short, simple motion works best for a first try — a cat dancing, a flag waving, a cup steaming. Busy scenes with five things moving at once tend to come back muddled.

Change the speed and fix a single frame

The part that saves the most time is frame editing. Once the GIF exists, you can ask ChatGPT to change one frame instead of generating the whole loop again. Say the cat’s paw looks wrong in the third frame, or the background shifts between two frames, and the model edits that frame while the rest of the loop stays as it is. That turns a near-miss into a finished GIF in one more message, where the old way meant rolling the dice on a full regeneration and losing the version you liked. Speed works the same way, as a follow-up in the same chat: say the loop plays too fast or too slow and it adjusts the timing. Keep refining in the same thread, so each edit builds on the GIF you already have instead of starting from a blank page.

Turn a still photo into a moving image

You do not have to start from a text prompt. Upload a still photo and tell ChatGPT what should move in it. A portrait can nod or smile, a street photo can gain passing traffic, a product shot can spin gently for a listing. The photo gives the model the subject, the lighting, and the framing; your words supply only the motion.

Be specific about how much should move: “Make the dog’s tail wag” tells the model to leave the rest of the frame alone, while “animate this photo” leaves everything up for grabs.

If ChatGPT hands you a still image instead

This is the part most early guides skip, and it is the one most readers will hit. Multiple independent account checks between 5 and 8 November found GIF controls missing: one tester got two still images and no GIF options, another asked Codex for a GIF and received a single-frame PNG, and the Help Center still described GIF creation as “rolling out” as of 7–8 November. Your account may simply not have the feature yet. That is a rollout gap, not something you did wrong.

OpenAI’s own launch examples point to the workaround: a sprite sheet. Ask ChatGPT for a single image holding six sequential frames in a 3-column by 2-row grid, with no borders or gutters between the frames, read left to right and then top to bottom. Stitch that sheet into a loop with FFmpeg, which is free — the sheet holds the same six frames a GIF would, and the stitching turns them into a loop you can share anywhere. Once your account gains native GIF output, the same prompts work there.

Save a photo so ChatGPT remembers it next time

Images 2.6 also lets you save photos as references for later chats. Save a photo of yourself, a pet, a product, or a character once, and you can call on it again in a new chat without uploading it again. The saved reference keeps the subject consistent across images you make days apart — the same dog in five different scenes, or the same product shot in a series of posts.

This is the quiet feature in the release, and for anyone making a run of images it may matter more than GIFs. A single saved reference replaces a fresh upload and a fresh description every time, and it removes the small differences that creep in when you describe the same subject from scratch in each chat. Our image-prompts guide covers how to describe a subject well in the first place — a saved reference does the same job with the photo itself.

Still images got faster too

GIFs are the headline, but the still-image side of Images 2.6 got a speed lift as well. OpenAI says generation is about 50% faster than before. That is the company’s own claim — no independent test has confirmed the figure — but the direction is clear: shorter waits between prompt and picture, which matters most when you are iterating on a design.

The speed claim also shows up in the API models. GPT-Image-2.6 Flare, the speed model, takes as little as 6 seconds at low quality, according to OpenAI. Again, treat that as the vendor’s number, not a measured result.

Flare vs Sunburst: which model does what

Images 2.6 ships as two API models with the same names as the 2.5 pair they replace. The 2.5 versions arrived on 8 September 2026; the 2.6 pair takes their place at the same pricing — US$8 per million image-input tokens, US$30 per million image-output tokens, and US$5 per million text-input tokens.

Model	Built for	When to pick it
GPT-Image-2.6 Flare	Speed	Quick drafts, fast iterations, and GIF work where turnaround matters more than fine detail
GPT-Image-2.6 Sunburst	Detail and consistency	Finished images where a face, a product, or a scene has to stay true across versions

The choice matters most if you build with the API, where you pick the model directly. The API also exposes GIF controls that the chat interface keeps hidden: output_format="gif", frame_count, and frame_duration_ms, with defaults of 9 frames at 200 ms each. Our Nano Banana guide walks through a similar speed-versus-quality trade-off in Google’s image model — the decision works the same way here.

What happens to GIFs that break the rules

Moving images get their own safety treatment. The Images 2.6 system card gives GIFs a dedicated section, and OpenAI tested its filters on a set of deliberately hard, harmful GIFs — the kind designed to slip past detection. The filters caught 77.7% of that test set. That figure comes from OpenAI’s own test — a self-reported score, not an independent audit. It also tells you something useful: nearly a quarter of deliberately tricky harmful GIFs got through the company’s own test, which is why some prompts will be refused even when a still-image version of the same idea would pass.

Where it still falls short

Three gaps are worth knowing before you plan work around this release. First, the rollout is staged and OpenAI gives no completion date, so two people on the same plan can have different features on the same day. Second, GIF access on the Free plan is not confirmed: OpenAI’s pricing page lists Free as “Limited image creation”, and the announcement’s “all subscription tiers” line describes Images 2.6 as a whole, not every feature inside it. Codex image generation is not on the Free plan, and the API has no free image route. Third, OpenAI publishes no per-plan GIF allowances or quotas, and its own docs disagree on the API frame-count range — one page says 2–64 frames, the API reference says 2–16 — so this guide states no maximum at all rather than picking a number that might be wrong.

My honest take: judge this release on the GIF workflow, not the speed claims. A 50% speedup is nice, but the thing that changes how people use ChatGPT images is asking for a moving image in plain words and fixing one frame instead of starting over. The speed numbers are vendor claims; the GIF workflow is something you can try in the next five minutes and feel for yourself.

FAQ

Next step: open ChatGPT, upload a photo of your pet, and ask for the cat-dancing treatment — one moving subject, one clear action. If a GIF comes back, change its speed once and fix one frame; if a still image comes back, ask for the six-frame sprite sheet and stitch it into a loop. Either way you will know within five minutes whether your account has GIFs yet, and what your fallback looks like until it does.