GPT Image 2.6 vs GPT Image 2.5 prompts: OpenAI kept its guide, creators wrote less

Summary

If you have GPT Image 2.5 prompts that work, keep them for GPT Image 2.6. Rewrite only for the three things that are new: GIFs, circled edits and the complexity setting.

The reason is simple: OpenAI did not change its own advice. The 2.6 version of its image prompting guide keeps all 24 example prompts from the 2.5 version, word for word, and adds three: a circled edit, a looping GIF and a complexity comparison.

Creators changed more than OpenAI did. We compared 238 GPT Image 2.5 prompts and 67 GPT Image 2.6 prompts that creators posted on X, each kept verbatim with its output:

• Shorter. In each model’s launch week, English prompts for still images went from a median of 203 words to 98, and from 7 to 2 lines.
• Fewer templates. 14% of 2.5 launch-week prompts left a blank to fill in, like [STRUCTURE NAME]. None of the 2.6 prompts did. Labeled sections such as Lighting: fell from 30% to 13%.
• Half are GIFs. 31 of the 67 creator prompts for 2.6 make an animated GIF, and those have a median of 68 characters. 大阪城を建築する流れをGIFにして (17 characters, build Osaka Castle as a GIF) is a whole prompt.

Treat the creator numbers as an early read: the 2.6 sample is two days old and small. The guide comparison is not; it is a plain text diff.

Our picks

1. Keep your 2.5 still-image prompts. Nothing in OpenAI’s guidance asks you to rewrite them. All 24 old examples, from the 1969 festival crowd to the infographics, carry over unchanged.
2. If you trim, trim toward OpenAI’s length. Its examples have a median of 71 words in 4 lines. The typical 2.5 creator template in launch week ran 203 words, nearly three times that. A long template is not wrong, but the extra length is your choice, not the model’s requirement.
3. Use one line for a GIF. Half of the 2.6 creator prompts are GIFs and the median is 68 characters. Our GIF measurements show what to add when one line is not enough: a frame grid, how the loop ends, the caption text.
4. Quote the exact words you want in the image. 27% of OpenAI’s examples put the required text in quotes. Only 7% of the 2.6 English still-image prompts did, and it is the cheapest way to control text rendering.
5. Copy the three new examples for the new features. Circled edits, looping GIFs and the complexity setting are the only places where OpenAI wrote new guidance. Each example is on our GPT Image 2.6 page with its output.

Which kind of prompt are you about to write: a still image you already have a template for, or one of the new features? Only the second needs a new prompt.

How we tested

1. The prompts. Every GPT Image 2.5 and 2.6 entry in the open collection awesome-image-video-prompts at commit 590b532: 359 prompts in all. 50 come from OpenAI’s prompting guide (24 for 2.5, 26 for 2.6), 305 from creators’ posts on X (238 for 2.5, 67 for 2.6), and 4 from other collections, which we left out.
2. Launch week. X post IDs carry their creation time, so we read the posting day of every creator prompt and counted days from each release: September 8 for GPT Image 2.5 and November 4 for 2.6. Launch week is days 0 to 6. That gives 69 launch-week prompts for 2.5 and all 67 for 2.6, which were posted in the first three days.
3. A fair subset. The two sets differ in language and in GIFs (2.5 cannot make them), so the headline comparison uses English prompts for still images only: 66 for 2.5 and 15 for 2.6.
4. Features. For each prompt we counted characters, words and lines, and tested a few patterns: labeled sections (two or more lines like Style:), lists, fill-in blanks, quoted text, an aspect ratio, camera terms, negative instructions (no, avoid, without), and whether the prompt starts from an uploaded image. The exact patterns are in the script.
5. The guide. We compared the 2.5 and 2.6 guide examples as normalized text. Our 2.6 set has 26 examples because the collection skipped one as a duplicate; we checked the live guide and that 24th example is there unchanged too.

Results

OpenAI’s guide, 2.5 to 2.6

	2.5 guide	2.6 guide
Examples	24	27 (24 carried over word for word, 3 new)
Median length	80 words, 4.5 lines	71 words, 4 lines (our 26)
Quote the text to render	29%	27%
Start from an uploaded image	50%	46%

Creators, launch week, English still images

	GPT Image 2.5 (66)	GPT Image 2.6 (15)
Median length	203 words, 1,458 characters	98 words, 727 characters
Median lines	7.5	2
Labeled sections	30%	13%
Lists	21%	7%
Fill-in blanks	14%	0%
Quoted text	26%	7%
Aspect ratio named	53%	27%
Negative instructions	62%	40%
Start from an uploaded image	44%	20%
Camera terms	20%	20%

All 67 creator prompts for 2.6

31 are GIFs, with a median of 68 characters and 1 line. The 36 still-image prompts have a median of 274 characters in any language. 42 of the 67 are in Japanese or Chinese, against 6 of 238 for 2.5; see Limitations before reading anything into that.

Is the drop just because launch-week posts are less polished? We checked by likes. Among 2.5 English still-image prompts with 20 to 100 likes, the median is 2,045 characters and 11 lines. No 2.6 prompt in any likes band comes close: the three in the 20 to 100 band have a median of 727 characters on one line. Popularity does not explain the gap.

A typical 2.5 launch-week template, the architectural sketchbook study, runs 13 lines and opens with Create a stunning 4:5 vertical fine-art sketch illustration of [STRUCTURE NAME]. A typical 2.6 still prompt, the cyan poetry portrait, is one dense paragraph that names the subject, the light and the projected text, with no blanks and no headings.

What to ignore

• The language shift. Two thirds of the 2.6 prompts are Japanese or Chinese because the 2.6 collection searched in those languages and read posts in the original. Earlier, X showed the collector auto-translated posts. It says nothing about who uses which model.
• Likes. The 2.5 set was picked from Top results weeks after release; the 2.6 set includes Latest results from the first days. Comparing likes across them compares two search settings.
• Shorter means better. We measured how people write, not which prompts give better images. OpenAI’s unchanged guide is the stronger evidence on what works.

Limitations

• The 2.6 English still-image sample is 15 prompts. One more week of posts could move the medians a lot. We will rerun the script on November 20 and update this post.
• The two sets were collected differently: 2.5 mostly from Top searches and open collections in late October, 2.6 from Top and Latest searches with a low likes bar in the first days after launch.
• Launch week for 2.5 covers seven days and for 2.6 three. Early posts tend to be quick tries, so part of the gap may be timing.
• The features are text patterns. A ratio like 16:9 inside a sentence counts; a prompt that says “make it wide” does not.

Reproduce it

The script is gpt-image-2-5-vs-2-6-prompts-2026-11.py. Clone the public collection, check out commit 590b532 and run the script on it; it needs only Python and PyYAML. It writes gpt-image-2-5-vs-2-6-prompts-2026-11.json: one row per prompt with its features, posting day and a link to the source file, plus every group summary in this post. Prompt text is not copied into the data; follow the link in each row.

Sources

1. OpenAI: Introducing Images 2.6 · official
2. OpenAI image prompting guide · official
3. awesome-image-video-prompts (the prompt collection we measured) · blog
4. AgentCraft: GPT Image 2.6 prompts, each with its output · blog