Chrome Extension
OpenAI released ChatGPT Images 2.6 on Nov 4, 2026, and the headline feature is GIF creation.
You describe a loop or upload a picture, and ChatGPT returns a short animated GIF. I read the official docs, the hands on tests and the leaderboards, then built 15 prompts that push each feature.
This review covers how the GIF tool works, what testers found, the SVG option that only the API documents, and where the output fits in marketing, branding and advertising.
Key Takeaways
What ChatGPT Images 2.6 Adds
Images 2.6 builds on the Images 2.5 models and adds three things you can use today: GIF creation, saved reference images and two API models. OpenAI also says generation is 50 percent faster, though it names no baseline. The sections below separate what OpenAI documents from what testers observed.
Release and Access
OpenAI announced the update on Nov 4, 2026. GIF creation runs in ChatGPT and Codex on web, desktop and mobile. OpenAI says it reaches all subscription tiers, and Tom's Guide reports that free users get it too.
The rollout looks staged. Notebookcheck said on Nov 8 that access was still spreading, and one Hacker News commenter reported that the Codex command line tool could not make GIFs on day one. Check your own account before you plan a campaign around it.
GIFs, Reference Images and Speed
You ask for a looping animation in plain language, or you upload a sketch or photo and ask ChatGPT to animate it. OpenAI's community announcement also lists adjustable playback speed and per frame editing. No hands on tester I found has used either control, so I treat them as reported and untested.
Reference images work through an @ mention. You save a photo under a name such as Chester, then write "@Chester on the moon" in a later chat. TechRadar found that ChatGPT stores the original photo and not the edited versions. OpenAI has not published how many references you can save.
Flare and Sunburst
The API offers two models, gpt-image-2.6-flare and gpt-image-2.6-sunburst. Flare targets speed and Sunburst targets quality, text and consistency. each::labs, a vendor that sells API access, ran one brief five times and suggests Sunburst for text and count heavy work and Flare for iteration.
Both models cost $30 per 1 million image output tokens. Kingy and CellCog estimate output only costs for a 1024 by 1024 image at $0.0247 on low quality, $0.0435 on medium, $0.0527 on high and $0.2107 on max. Treat those as estimates, because OpenAI publishes no flat per image price.
How the GIF Feature Works
The GIF tool is an image model that draws a set of related frames and strings them together. That design explains both its strengths and its limits.
A Flipbook and Not a Video
The Decoder describes OpenAI's cat demo as several slightly varied images strung together like a flipbook. The Neuron saw the same thing: its test GIF came from an 8 pose sprite sheet. Webiano argues that this suits short loops and fits poorly with camera moves, complex physics, several characters interacting, or anything that does not return to its starting pose.
So I would not compare it with a text to video model. It is closer to a fast way to make a small looping animation from one idea.
Limits You Can Plan Around
OpenAI has not published GIF limits for the ChatGPT app. The API guide does publish its own limits, and testers filled in what they saw in both places.
| Setting | What testers saw (ChatGPT and API) | API limit (OpenAI's guide) |
|---|---|---|
| Frames | 6 to 12 | 2 to 16, default 9 |
| Frame delay | 150 to 200 ms | 20 to 2000 ms, default 200 |
| Loop length | 0.9 to 2.2 seconds | About 1.8 seconds at the defaults |
| Size | 358 to 832 px wide in different tests | Multiples of 16, with a total pixel budget OpenAI does not publish |
| Looping | Ask for it in the prompt | No loop setting, so ask in the prompt |
Some third party sites list up to 64 frames. OpenAI's own guide says 16, so test against that number. The guide also says the API has no streaming for GIFs, and some frame sizes get rejected when they exceed the pixel budget.
Why Nine Frames Can Look Stepped
This is my own math and not a tester finding. Nine frames at 200 ms each play at about 5 frames per second, while practitioner guides put common GIF frame rates at 10 to 20. If your default output looks choppy, raise frame_count or lower frame_duration_ms in the API, or ask for more frames in the prompt.
What Testers Found
Only a handful of people have published real hands on GIF tests. Most coverage restates the launch post, and the results below come from the few that ran prompts.
Results That Worked
TechRadar's first cat GIF, a Halloween scene, was acceptable but not lively. After the author asked ChatGPT to make it more active and creative, the cat raised a paw like casting a spell next to a floating glowing pumpkin. The author called the movement realistic and much better.
Analytics Vidhya asked for a clay style copper teapot with steam forming a heart. The result was a 12 frame loop of 2.16 seconds that the author called smooth. A DEV Community tester got a waving robot at 832 pixels in 6 frames, and a line art bird with a reference image held its line style across blink, hop and heart frames.
Where It Broke
Kingy tested a rocking water bottle through the API and the ChatGPT web app. Both runs had uneven motion and a visible jump back to the start pose, and both added a mark to the bottle's badge that nobody requested. The web version also came out at 358 pixels instead of the 512 requested.
The Neuron saw a dancing cat stay recognizable while the strawberry on its head changed shape between frames. On Hacker News, one commenter called a chess animation splotchy and unusable, another saw a windmill door flicker, and a third said splotches got worse when they regenerated from earlier outputs. One user's guardrail refusal went away after a rewording, which tells me the safety checks are sensitive to phrasing.
How Long It Took
Timings are single observations. Kingy's API run took 17.6 seconds and DEV's runs took 20 to 25 seconds. The Neuron's cat took 2 minutes and 21 seconds inside ChatGPT, and Kingy's web runs took 38 to 44 seconds.
Sharing from a phone also gave one commenter a ChatGPT link that showed a static image. TechRadar did confirm a share link, a download option and posting to X, LinkedIn and Reddit.
The SVG Angle
SVG is where Images 2.6 gets interesting for brand work, because a vector logo or icon stays sharp at any size. OpenAI documents it, but the details matter.
What OpenAI Documents
Both models accept output_format set to svg on the image generations endpoint. OpenAI's image generation guide says the API first makes a reference image, then a second model writes the SVG, and you pay for both steps. You can request one image per call, and the edits endpoint cannot output SVG.
The result is static. It carries no scripts, no external resources and no embedded raster image, so any animation is your job. The guide shows a bakery logo from both models and recommends SVG for icons and illustrations.
I found no OpenAI statement that the ChatGPT app exports SVG. A few sites claim it, others say API only, and none of the claims I saw came from a test in the app. Plan as if SVG is an API feature.
The One Reliability Test
CellCog ran 48 SVG requests on Nov 4 and 5. Nine succeeded, which is 19 percent, and the rest returned HTTP 502 errors. Text heavy prompts succeeded 3 of 41 times.
The files that did come through weighed 0.3 to 14 KB, cost $0.03 to $0.23 each, and kept their text as live text. CellCog sells a Recraft based tool and preferred Recraft overall, and launch week traffic may explain the errors. I read this as a warning to build retries into any SVG workflow, not as a verdict on the model.
GIF, SVG or MP4
Each format wins a different job.
| Format | Best for | Watch out for |
|---|---|---|
| GIF from Images 2.6 | Characters, memes, product shots, email animation | 256 colors per frame and hard edged transparency. Outlook 2007 to 2019 shows only the first frame, and Litmus advises keeping email GIFs under 1 MB |
| SVG from the API | Logos, icons, loaders, infographics | Static output. Gmail now converts SVG to a flat PNG, so animation is lost there, according to Can I Email's September 2026 test |
| MP4 or WebM | Photographic motion, social posts | web.dev shows a 3.7 MB GIF shrinking to 551 KB as MP4 and 341 KB as WebM. HTML video does not play in Gmail or Outlook on Windows |
The rule I take from this: use GIF where email support matters, SVG where sharpness and file size matter, and video where the motion is photographic and the platform allows it. Sources for the table: Litmus, Can I Email for SVG, Can I Email for video and web.dev.
A Workflow That Combines Them
You have two routes to an animated vector. In the first, you generate an SVG with the API and add CSS or SMIL animation yourself. In the second, you generate a raster image in ChatGPT, trace it to SVG with a tool such as Vectorizer.AI or vtracer, and then animate with CSS.
Tracing has limits. A vendor test found that traced photos posterize and gradients lose detail, so tracing suits flat logos and icons and fails on photo style art. Whichever route you pick, add a prefers-reduced-motion rule and export a GIF or MP4 fallback for email and social. I found no published comparison between the API's SVG and a traced version of the same prompt, which makes it a good first experiment.
Images 2.6 Review: Scores, Speed and Price
This is my verdict from the published evidence, since I did not generate outputs myself. The short version: strong on benchmark quality, good for simple loops, risky for text and logos, and unproven for SVG.
Leaderboards
On Artificial Analysis, the API model 2.6 Sunburst ranks first at 1259 Elo. 2.6 Flare scores 1228, 2.5 Sunburst 1227 and Google's Nano Banana 2.1 scores 1176. The scores are for API models at max quality, and the page lists GPT 2.6 at max quality at $210.72 per 1,000 images against $33.60 for Nano Banana 2.1, which is not a like for like comparison.
DesignArena also ranks 2.6 Sunburst first at max quality, 1385 against 1380 for 2.5 Sunburst, as of Nov 8. Arena's preliminary Nov 4 snapshot, reported by Kingy, put 2.6 Sunburst second at 1409 behind 2.5 Sunburst at 1427. Kingy warns that settings differ across models, so read these as direction and not as proof.
Speed
OpenAI claims generation is 50 percent faster and gives no baseline. CellCog ran 80 test runs on one laptop and found 2.6 Flare about twice as fast as 2.5 Flare, with high quality dropping from 17.2 to 7.9 seconds. It also found 2.6 Sunburst slower than 2.5 Sunburst on every prompt.
So the speed claim holds for Flare and not for Sunburst. If you iterate on GIF ideas, start with Flare and switch to Sunburst for the final version.
Strengths and Weak Spots
Here is where the evidence points.
Rights, Provenance and Safety
OpenAI's system card page says ChatGPT images keep C2PA metadata and carry an invisible SynthID watermark. I could not confirm that GIF or SVG outputs carry either, and platforms can strip metadata when you convert files. OpenAI's terms of use say you own your output and can use it commercially if you follow the terms, so read them before a paid campaign.
Real people need care. OpenAI's usage policies bar photorealistic images of real people without consent where they could mislead. Notebookcheck points out that saved references plus animation could place real people in made up scenes, and OpenAI has not said how it guards against that. I found no OpenAI rule that names brand logos or trademarks for images, so use your own marks and fictional brands in tests.
Where GIFs Fit in Marketing, Branding and Advertising
A short looping GIF works when the motion carries one idea. It fails when the motion is decoration.
Marketing
Email is the strongest fit. Mailchimp calls a GIF a stand in for video that plays inside the inbox, and Litmus lists small motion on calls to action, urgency banners and loyalty tiers as working uses. Social teasers and landing page loops also suit a 1 to 2 second flipbook.
Motion needs a purpose. The Nielsen Norman Group tested 14 ecommerce emails in 2020, and viewers rated the animated versions lower on trust (3.90 against 4.34) and net sentiment (minus 0.36 against 0.55). The sample was small, but it supports a simple rule: animate the one thing you want noticed.
Branding
Reference images make a mascot set realistic. You can save one character, then ask for celebrating, thinking and waving versions that keep the same face and colors. GIPHY's sticker guide asks for looping, transparent GIFs, with a maximum of 15 seconds and 6 seconds advised.
Keep exact logos out of generated frames. A logo that drifts a few pixels from frame to frame looks like an error, so animate your mascot and place the real logo in your design tool afterward.
Advertising
Google Ads accepts GIFs for uploaded display ads with a 150 KB limit, 30 seconds and under 5 frames per second. That frame rate rule favors a slow flipbook, which is the style Images 2.6 produces. The file size is the harder limit, because one tester measured a 6 frame, 508 pixel GIF at about 452 KB. Expect to compress it before it fits a 150 KB slot, and note that the IAB guide caps animation at 15 seconds.
Ecommerce and Other Uses
Shopify supports animated GIF and WebP in product media, with images under 20 MB. SaaS teams can use a short loop for an onboarding tip or an empty state. Support teams should be careful, because HelpDocs moved its knowledge base from GIFs to video, and Splunk's style guide caps GIFs at 20 seconds and asks writers to supplement them with text.
Platform Limits at a Glance
| Where | What I verified |
|---|---|
| Keep GIFs under 1 MB. Outlook 2007 to 2019 shows the first frame only, while Microsoft 365, 2021 and 2024 play GIFs by default | |
| X | 15 MB, 1280 by 1080 recommended, 350 frames. HubSpot lists 5 MB on mobile, so test both |
| No native GIF feed post. Convert to MP4 or use a Stories sticker | |
| Slack | Custom emoji under 128 KB is recommended |
| GIPHY | Stickers need looping, transparent GIFs of 15 seconds or less |
| Tenor | The API shut down on June 30, 2026, so do not build on it |
Accessibility
WCAG 2.2.2 requires a way to pause or stop motion that autoplays for more than 5 seconds, and 2.3.1 limits flashing to three per second. Short loops, a static fallback and a prefers-reduced-motion rule on web pages handle most of it. That rule does not work in Gmail or Outlook on Windows, so email needs a calm first frame.
15 Test Prompts for ChatGPT Images 2.6
I wrote these prompts to push each feature, one use case at a time. I have not run them, so treat my "watch for" notes as what the evidence says to check and not as results. Every prompt uses a fictional or placeholder brand, so swap in your own mascot, product and colors.
| Feature | Prompts | Use cases |
|---|---|---|
| Text to GIF | 1, 2 | Email header, display banner |
| Photo or sketch to GIF | 3, 4 | Ecommerce product, brand mascot |
| Reference images | 5, 6 | Mascot reaction set, campaign scenes |
| Text inside motion | 7, 8 | Offer badge, software onboarding tip |
| Loop quality and pace | 9, 10 | Seamless loop, speed |
| Stress tests | 11, 12 | Camera orbit, transparent sticker |
| API GIF and SVG | 13, 14, 15 | Developer workflow, logo, icon set |
How to Run Them
Replace every [bracket] before you send a prompt, and run each one twice in a fresh chat. Then score every result on the same five checks.
Prompt 1: Email Header Loop (Text to GIF, Marketing)
This tests the simplest job: one object, one small motion, and a clean loop for an email.
Create a looping animated GIF for an email header, wide 2:1 format. Subject: a [product] on a clean [color] background. Motion: one small movement only, such as [steam rising / the lid lifting and settling / a light sweep across the label]. Camera: fixed. Background: fixed. Start and end on the same pose so the loop is seamless. Style: [flat illustration / soft studio photo]. No text and no logos.
Watch for: a jump at the loop point and unrequested marks on the product. Kingy's tests showed both.
Prompt 2: Display Banner Loop (Text to GIF, Advertising)
This tests a slow, low frame count loop that suits ad rules, and it shows how big the file comes out.
Create a looping animated GIF for a display ad, wide 4:1 format. Scene: a [product] on the left and empty space on the right where I will add text later. Motion: the product bobs up and down slightly while a highlight sweeps across it once. Use 6 frames at a slow pace of about 3 frames per second. Fixed camera, flat [brand color] background, last frame matches the first. No text and no logos.
Watch for: file size. Google's uploaded display GIF limit is 150 KB, and one tester's 6 frame GIF weighed about 452 KB.
Prompt 3: Product Shimmer (Photo to GIF, Ecommerce)
Upload a clean product photo first. This tests whether the product stays identical while one thing moves.
Use the attached product photo as the first frame. Make a looping GIF where the only change is a soft light reflection that slides across the product once. Keep the product shape, color, label text and background exactly as they are in the photo. Fixed camera. 8 frames. The last frame matches the first.
Watch for: label text that shifts or blurs, and shape drift between frames.
Prompt 4: Mascot Sketch Wave (Sketch to GIF, Branding)
Upload a simple mascot sketch. OpenAI's own examples include "make the character wave," so this tests the basic claim.
Animate the attached mascot sketch into a looping GIF. Action: the mascot waves its right hand twice, then returns to the starting pose. Keep the same line style, colors and proportions on every frame. Plain white background, fixed camera, 8 frames.
Watch for: the wrong hand. One tester reported that the wave direction did not match the prompt.
Prompt 5: Mascot Reaction Set (Reference Images, Branding)
Save your mascot image as a reference named Mascot first. The menu wording may differ by account, so look for the reference image option in your settings or in the image tool. Then run this prompt five times, changing only the action.
Make a looping GIF of @Mascot [celebrating / thinking / giving a thumbs up / falling asleep / waving goodbye]. Keep the same face, colors and outfit as the saved reference. Use one small action, then return to the starting pose. Plain [brand color] background, fixed camera, no text.
Watch for: identity drift across the five GIFs. TechRadar saw the reference work, with some questionable details.
Prompt 6: Campaign Scenes (Reference Images, Marketing)
This tests whether the same character survives a change of scene while the scene itself stays still.
Make a looping GIF of @Mascot on [a launch day billboard / a beach / a snowy mountain]. Keep the mascot identical to the saved reference. Only the mascot moves: [one small action]. The scene stays still. Seamless loop, fixed camera.
Watch for: background elements that move when they should not, and odd logic. TechRadar's moon scene raised questions about the butterfly and the helmet.
Prompt 7: Offer Badge (Text Inside Motion, Advertising)
Text is the classic weak spot, so this prompt demands exact words on every frame.
Create a looping GIF of a round badge that reads "SAVE 20" in bold white letters on a [brand color] circle. Motion: the badge tilts left and right by 5 degrees while small sparkles appear and fade around it. The text must read exactly "SAVE 20" in every frame. Fixed camera, plain background, seamless loop, no other words.
Watch for: one wrong letter in a single frame. Open the GIF in a viewer that lets you step through frames.
Prompt 8: Onboarding Tip (Text Inside Motion, Software)
This tests a small interface scene for a help page or an onboarding email.
Create a looping GIF for a software onboarding tip. Show a simple app window with a button labeled "Export" and a cursor that moves to the button, clicks it, then returns to its start. Use the exact label "Export" in every frame and no other text. Flat vector style with [brand color] accents, fixed camera, 8 frames.
Watch for: invented menu text and a cursor that teleports. For real product walkthroughs, record video instead, since HelpDocs moved its docs from GIFs to video.
Prompt 9: Seamless Loop (Loop Quality, Marketing)
This tests the loop on a soft, organic motion that is easy to get wrong.
Create a seamless looping GIF of a ceramic mug with steam that curls into the shape of a coffee bean and fades. Frame 1 and the last frame must look identical. Soft clay stop motion style, warm light, fixed camera, plain background. Use between 10 and 12 frames.
Watch for: the seam and for splotchy texture. Analytics Vidhya got a smooth 12 frame loop from a similar idea.
Prompt 10: Pace Test (Speed, Social)
Run the same prompt three times with a different pace word, then use the speed control in the viewer, if you see one, on the best result.
Create a looping GIF of a paper boat rocking on gentle water. Pace: [slow and calm / medium / quick and playful]. Same boat, same water, same camera in every version. Seamless loop.
Watch for: whether the pace word changes the motion or only the mood. OpenAI lists adjustable playback speed, but no tester has reported using it yet.
Prompt 11: Camera Orbit Stress Test (Limits, Advertising)
I expect this one to struggle, and that is the point. It shows you where the flipbook approach stops.
Create a looping GIF where the camera orbits a full 360 degrees around a [product] on a turntable. The product must stay identical from every angle. Plain background, 12 frames, seamless loop, no text.
Watch for: a product that changes shape between angles and a loop that does not close. Webiano says camera moves are a weak fit.
Prompt 12: Transparent Sticker (Limits, Branding)
GIPHY stickers need looping, transparent GIFs. GIF only supports one fully transparent color, so edges can look rough.
Create a looping sticker GIF of [mascot or character] giving a thumbs up, on a transparent background. Thick clean outline, flat colors, no shadow and no glow. Small movement only, seamless loop, 8 frames.
Watch for: halos around the edges. Place the result on a dark background and a light background to check. One Hacker News tester said a transparent background helped an eagle test, but that is a single report.
Prompt 13: API GIF Call (Developers)
This JSON body uses the parameters from OpenAI's guide. I built it from the guide and have not run it. Send it as the body of a POST request to /v1/images/generations with your API key, and follow the guide to save the response.
{
"model": "gpt-image-2.6-flare",
"prompt": "A looping animation of a paper boat rocking on calm water. Fixed camera. The last frame matches the first.",
"output_format": "gif",
"frame_count": 12,
"frame_duration_ms": 100,
"size": "512x512"
}Watch for: 12 frames at 100 ms gives a 1.2 second loop at 10 frames per second, so compare it with a run at the defaults of 9 frames at 200 ms. If the API rejects a size, lower it, because the pixel budget is unpublished.
Prompt 14: API SVG Logo Mark (Branding)
This body asks for a static vector logo. It has no text on purpose, because the one reliability test I found failed most text heavy requests.
{
"model": "gpt-image-2.6-sunburst",
"prompt": "A flat logo mark for a fictional bakery called Northwind Bakery: a wheat stalk inside a circle. Two colors, no gradients, no text.",
"output_format": "svg",
"n": 1
}Watch for: the success rate. Run it 10 times and log how many calls return a file, then open the file and check that it contains vector shapes and no embedded image.
Prompt 15: API SVG Icon Set (Branding and Product)
This body tests consistency across several icons in one file.
{
"model": "gpt-image-2.6-flare",
"prompt": "A set of 6 flat line icons for a coffee shop app: cup, bean, receipt, map pin, gift card and bell. Same stroke width, one color, no text, no gradients.",
"output_format": "svg",
"n": 1
}Watch for: uneven stroke widths and missing icons. If the set works, add a short CSS animation to one icon and keep the others still.
Frequently Asked Questions (FAQs)
Is ChatGPT Images 2.6 free?
Possibly, but check your account. OpenAI says GIF creation reaches all subscription tiers, and Tom's Guide and TechRadar report that free users get it. Notebookcheck said on Nov 8 that the rollout was still staged, so you may not see it yet. API use is billed per token.
How long can a ChatGPT GIF be?
OpenAI has not published a limit for the ChatGPT app. Testers saw 6 to 12 frames and loops of about 1 to 2 seconds. The API accepts 2 to 16 frames at 20 to 2000 ms each, which would stretch to 32 seconds on paper, but that would play like a slideshow.
Can ChatGPT Images 2.6 make SVG files?
Yes, through the API. Both models accept output_format set to svg, and the output is a static vector file. I found no OpenAI statement that the ChatGPT app exports SVG, and one launch week test got only 9 of 48 requests through.
Should I use Flare or Sunburst?
Start with Flare when you iterate on ideas, because CellCog measured it at about twice the speed of 2.5 Flare. Switch to Sunburst for final work with text or exact counts. Sunburst tops the Artificial Analysis and DesignArena image leaderboards, and both models cost the same per token.
Do Images 2.6 GIFs work in email?
Mostly yes. Gmail and Apple Mail animate GIFs, while Outlook 2007 to 2019 shows only the first frame and Microsoft 365, 2021 and 2024 play them by default. Keep each file under 1 MB and design a first frame that works alone.
Is Images 2.6 a video model?
No. It draws a small set of related frames and strings them into a loop, so it fits short, simple motion. For camera moves or clips longer than a couple of seconds, use a text to video model such as Google's Veo, which makes 4, 6 or 8 second clips at 24 frames per second with audio.
Final Thoughts
ChatGPT Images 2.6 is a fast way to make short looping animations, and its benchmark scores are among the best published for image generation. The GIF results are mixed, with smooth loops from simple ideas and flicker, jumps and text slips from harder ones, and SVG is documented for the API but unproven in practice.
Your next step is to run prompts 1, 3 and 5 this week with your own product or mascot, and score each result on the five checks. Then run prompt 14 ten times before you build SVG into any workflow. A day of testing will tell you more about your brand than any review, including this one.
