ChatGPT GIF prompts for GPT Image 2.6: what 93 collected prompts show

Summary

If you want a GIF from ChatGPT Images 2.6, start with one plain line: make a GIF of what you want. Then add only the parts you need to control: a frame grid, how the last frame returns to the first, and any caption text.

The reason is simple. We collected 93 GPT Image 2.6 prompts in the three days after launch, each with the output its author posted, and measured all 43 animated GIFs frame by frame. Here is what they show:

• Short prompts work. 22 of the 42 creator GIF prompts are 50 characters or shorter. 大阪城を建築する流れをGIFにして (17 characters, build Osaka Castle as a GIF) came back as a 12-frame construction sequence.
• A GIF is a frame grid played back. OpenAI’s own launch prompts ask for a sprite sheet of 6 frames in 3 columns and 2 rows. 36 of our 43 GIFs have 4, 6, 9, 12 or 16 frames, and 22 have exactly 9.
• A frame count you ask for is kept. All five prompts with a checkable count got it, up to a 4×4 sheet of 16 frames.
• Asking for a loop helps. Without it, 10 of 28 GIFs jump visibly when they restart. With it, 2 of 15 do.
• Captions render. All eight prompts that asked for text on the image got legible text, in Chinese, Japanese and English.

Every prompt below links to its entry on our GPT Image 2.6 prompt page, with the output and the original post.

Our picks

What we would write, in order of how much it changes the result:

1. One line first. generate a gif of a pelican riding a bicycle (prompt) is the whole prompt, and it got a 9-frame GIF that cycles smoothly. Write more only when the first result misses.
2. Say how it ends. For a seamless loop, describe the last frame as the first one. The two cleanest loops in the set both do this: the Tanjiro pixel-art strike says First and last frame match so it loops seamlessly, and the ginger kitten ends with Frame 9: back to the kitten sitting small in the center, same as frame 1. OpenAI’s own monkey example does the same with then lowers it back to its starting pose.
3. Name the grid when timing matters. OpenAI’s launch prompts say Exactly 6 sequential frames arranged in a precise 3-column by 2-row grid and add no borders, no gutters. The pixel jungle outpost asks for a 4×4 spritesheet of 16 frames at about 200 milliseconds each, 3.2 seconds in all, and got 16 frames in 3.16 seconds.
4. Write captions word for word. 图片里有文字：真香！ (put the text 真香！ in the image) produced exactly that (prompt). The chat banner mixed Japanese, Latin letters and a time, and every character came out right.
5. Upload a character to keep it on model. 15 of the 42 creator GIFs started from an uploaded image. The most-liked post in the set, a sticker template with 895 likes, is built this way: lock the character from the upload, design one gag, add a 2 to 8 character caption, leave a slot for your own meme. A 12-sticker set goes further and gets 12 captioned stickers in a 4 by 3 overview from one upload.
6. Ask for length if you want more than a blink. Most GIFs run 1 to 2 seconds. Four of seven length requests were met, including 3 and 3.2 seconds; a request for 8 seconds came back at 4.

Which of these does your last GIF prompt already do? If it is a single line, try adding only item 2 and compare.

How we tested

1. The prompts. All 93 GPT Image 2.6 prompts on our prompt page as of November 6, 2026. 26 come from OpenAI’s image prompting guide, shown with the Sunburst output. 67 come from creators’ posts on X between November 4 and 6, each kept verbatim with the output the author posted. 43 entries are animated GIFs: 42 from creators and one from the guide.
2. How we found them. X searches for GPT Image 2.6 and Images 2.6, in Top and in Latest, in English, Japanese and Chinese, plus post links from other open collections. A reviewer model checked each prompt against its output, and we looked over every accepted entry. Most posts share a result without the prompt: in the last pass we opened 408 posts with an image or GIF, and 29 had a prompt that matched the output.
3. Prompt features. Length in characters (half the prompts are Chinese or Japanese, so word counts would mislead), language, whether the prompt uses an uploaded image, asks for a loop, gives a frame count or duration, or asks for on-image text.
4. GIF measurements. For each GIF we collapsed repeated frames and counted the distinct ones, read the total duration, and computed a loop jump: the pixel change from the last frame back to the first, divided by the median change between neighbouring frames. A jump of 1 means the restart looks like any other step; above 2 it is a visible snap.
5. Captions. We opened every GIF whose prompt asked for text and read the frames.

Results

Who wrote the 42 creator GIF prompts

	GIF prompts
Japanese	17
English	14
Chinese	11
Started from an uploaded image	15
50 characters or shorter	22
400 characters or longer	8

The median creator GIF prompt is 47 characters. The median OpenAI guide prompt is 447, and the median creator prompt for a still image is 727. People write long prompts for photos and posters, and one line for GIFs, and still get a GIF.

Frames

Distinct frames	GIFs
9	22
16	6
12	4
6	3
8	3
4	1
Other (17, 24, 24, 29)	4

36 of 43 counts are 4, 6, 9, 12 or 16, the cell counts of 2×2, 2×3, 3×3, 3×4 and 4×4 grids. OpenAI’s launch post explains why: its two GIF examples, made by OpenArt, begin with Create a single animation sprite sheet for a seamless looping GIF and ask for a fixed grid of equal panels. The model draws the frames as one sheet and plays the cells in order. With no grid named, 3×3 is the common result.

Requested counts were kept every time we could check: 9 frames twice, about 12 once, 6 to 8 once (it gave 8), and a 4×4 sheet of 16 once.

Duration

The median GIF runs 1.4 seconds and 36 of 43 run 2.1 seconds or less. Seven prompts asked for a length and four got it: 1.26 seconds asked, 1.25 measured; 3 seconds asked, 3.0 measured; 2 to 3 seconds asked, 3.0 measured; 3.2 seconds asked, 3.16 measured. The other three asked for 1.5 to 2, 2 to 3 and 8 seconds and came back at 1.3, 0.3 and 4.1.

The longest Chinese GIF prompt, a 2,139-character idiom riddle, asked for 3 seconds in four 0.75-second segments, one object lifting in each while everything else stays still. The GIF does exactly that: four separate lifts, evenly spaced, 3.0 seconds in total.

Loops

	GIFs	Visible jump at restart
Prompt asks for a loop	15	2
Prompt doesn’t mention one	28	10

Some of the 10 jumps are intended. The Osaka Castle build and the goodnight cat, which yawns, hugs a pillow and falls asleep under its caption, tell a story with an end. If you want a sticker that cycles forever, though, ask for the loop.

Captions

Eight prompts asked for text in the image: 真香！, ego be gone!, a Japanese chat banner with a time, おやすみにゃ～！, two meme captions the model wrote itself, the idiom riddle’s title, and a short line on each of the 12 stickers in the sticker set. All eight are legible and spelled as asked.

Likes

The median creator post has 7 likes. The top three have 895, 208 and 151: a sticker template, a 1,378-character photo prompt and a 36-character one-liner. Likes follow the size of the author’s audience, so read this as which formats people share, rather than which prompts make better images.

The 25 still prompts from OpenAI’s guide

They are longer and more explicit: median 506 characters, 12 edits and 13 generations. 7 of the 12 edit prompts list what must stay unchanged, such as face, pose, background and framing, before saying what to change. If you edit photos with GPT Image 2.6, that pattern is worth copying even though no creator GIF prompt needed it.

What to ignore

Prompt length as a score. A 17-character prompt and a 2,139-character prompt both produced the GIF their authors wanted. Length buys control over timing, frame order and the loop, so add it for those and stop there.

Frame rate requests. One prompt asked for 20 fps and about 60 frames. X re-encodes posted GIFs, so we can’t tell what rate the model produced, and you probably can’t either once you share it.

Limitations

• Three days of launch posts. Creators had had the model for one to three days. Japanese creators posted GIF prompts first, which is why Japanese leads the language count.
• Measured on X’s copy. X serves GIFs as MP4, and our previews were rebuilt from that MP4 at 12 fps. Durations and distinct frame counts survive this; exact frame timing does not.
• One GIF from the author’s reply. The orange tabby post shows a screen recording of ChatGPT. The author’s reply has a clean export of the GIF next to the line drawing, so we measured that copy, cut to one cycle.
• Posted outputs only. Authors share the results they liked. We see what worked at least once, and nothing about how many tries it took.

Reproduce it

Per-prompt numbers are in gpt-image-2-6-prompts-2026-11.json: source, mode, prompt language and length, likes, and for each GIF its distinct frames, duration and loop jump. Each row links to the prompt on our GPT Image 2.6 page, where you can copy it and see the output.

Sources

1. OpenAI: Introducing Images 2.6 (Animate your ideas with GIFs) · official
2. OpenAI image prompting guide: Generate a looping GIF · official
3. OpenAI API docs: gpt-image-2.6-sunburst · official
4. AgentCraft: GPT Image 2.6 prompts, each with its output · blog