ChatGPT Images 2.6: GIFs, SVG and What Marketers Must Check

OpenAI released ChatGPT Images 2.6 on 4 November 2026, adding animated GIF output, SVG export, saved reference images and two API models, GPT-Image-2.6 Flare and Sunburst. Marketers can now make short looping assets and vector files from a prompt. Provenance and likeness rules are less settled than the feature list, so treat the output as draft material until you have checked both.

Key Takeaways

• OpenAI's API changelog dates the launch of GPT-Image-2.6 Sunburst and Flare to 4 November 2026, with adjustable complexity, SVG and GIF output, and markup-guided edits.
• In ChatGPT, GIFs are built from a prompt, a photo or a sketch, and you can save a named reference image (a person, pet or character) and recall it with an @ mention. MIXED News reported on 7 November that GIF creation was still rolling out.
• The "50% faster" figure is OpenAI's own, with no stated baseline. Flare's "about six seconds" applies only to low quality.
• Low and medium quality images cost noticeably more per image than on Images 2.5, because they use more tokens. High and max cost the same.
• Whether GIF and SVG files carry C2PA metadata or a watermark is not confirmed by anything I could open. Do not rely on the file to disclose that it is AI-made.
• The API's SVG output is static, generation-only and limited to one file per request. The edit endpoint does not accept SVG.
• I could not find any source for an administrator preview window ending 18 November. Treat that date as unverified.

What OpenAI actually shipped

The primary source I could open is OpenAI's developer documentation. The changelog entry for 4 November lists the two models, complexity control, SVG and animated GIF output, and markup-guided editing for supported GPT-Image-2.5 and 2.6 models. OpenAI's own announcement page and ChatGPT release notes returned access errors for me, so the consumer-side details below come from secondary coverage, and I name it each time.

One conflict to flag: i-scoop reports a GIF range of 2 to 64 frames, while OpenAI's guide says 2 to 16. I trust the guide. Flare is the faster model and Sunburst is the one OpenAI positions for detail and consistency across edits at medium or high quality.

Speed, price and quality: what is verified

OpenAI says Images 2.6 generates images 50% faster and that Flare can finish in about six seconds at low quality. Kingy.ai, which summarizes OpenAI's documentation and labels its own commentary, points out that the baseline and workload are not specified, so the claim does not establish a 50% cut in elapsed time for your jobs.

On cost, the per-token rates are the same for both models: $5 per million text input tokens, $8 per million image input tokens and $30 per million image output tokens, per Kingy.ai's reading of OpenAI's pricing. What changed is how many tokens a low or medium image uses. Using OpenAI's cost calculator as reported on 7 November for a 1024 by 1024 output:

i-scoop reports the same pattern in tokens (low rises from 196 to 823). So if you ran cheap, low-quality drafts at volume, your bill can rise roughly fourfold per image. Kingy.ai also notes that no Batch rate is listed for 2.6, while 2.5 has one.

On quality, independent boards exist but are early. Kingy.ai reports Rapidata scores of 1316 for Sunburst and 1293 for Flare, and preliminary Arena scores at medium quality. These boards use different scales and cannot be pooled, and none tests GIF or SVG quality. Early users quoted by i-scoop described "splotchy" GIF frames and unspectacular SVG results, which is one source and anecdotal.

Rights, provenance and disclosure

This is the part most feature roundups skip.

Provenance. Kingy.ai, citing OpenAI's system card, says the company applies C2PA metadata and SynthID watermarking. But i-scoop could not verify that GIF and SVG files carry that metadata, and Webiano notes that OpenAI has not publicly confirmed a watermark survives conversion to GIF. Metadata is also easy to strip when a file is re-encoded or uploaded to a social platform. My read is that you should disclose in the caption or ad label, not rely on the file.

EU marking rules. Both i-scoop and Webiano say providers already on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking duty under Article 50 of the EU AI Act. I covered what the transparency rules ask of marketing teams in EU AI Act Article 50: What Marketers Disclose From Aug 2026 and the compliance table in EU AI Act Article 50 Is Live. Read those before publishing AI-made visuals to European audiences.

Likeness. Saved reference images accept people. MIXED News said it found no description of consent requirements or restrictions for people's photos, and its only example was a pet. Webiano also says it is unclear whether animating photos of real people is allowed. Until that is documented, only save and animate people who have given you written permission for commercial use, including staff, creators and students. The checklist in Before You Run AI Faces in Paid Ads applies directly.

Ownership and contracts. I did not open OpenAI's current terms on output ownership, so I make no claim on that point. Read them before using output in a client deliverable. Separately, i-scoop reports that Business, Enterprise, Edu and API accounts come with a data processing agreement while consumer plans do not. If you upload customer or student photos, that matters.

What this means for marketers and founders

My read, as opinion rather than tested result:

1. Test GIFs on low-stakes placements first. Email headers, blog illustrations and organic social posts are fine. Brand logos and product labels are where consistency errors cost you. MIXED News itself recommends checking text, faces and product details before using results in finished work.
2. Use SVG for simple marks, not for brand assets. The output is generation-only and static. Treat it as a starting sketch for a designer, not a final logo.
3. Re-price before you scale. If you built a pipeline on low-quality drafts, re-run the numbers using the table above. Medium quality is now roughly three times the 2.5 price per image.
4. Write a one-page likeness and disclosure rule. Name who may be saved as a reference image, who approves it, and how AI use is labeled. AI Generated Video Disclosure on Social Media is a good template to adapt.
5. Do not expect an SEO lift from the feature. Original images and clear alt text still matter for visual search, as I outlined in Nano Banana 2.1 in AI Mode. Generation speed does not change how a page is evaluated.
6. Keep a testing log. Before swapping tools, compare 20 of your own briefs across two models, the way I suggest in AI UGC Ads: A Testing Framework.

What we do not know yet

• Whether GIF and SVG outputs carry C2PA or SynthID marks.
• The consumer-side rules for animating real people's photos.
• The 18 November administrator preview date, which I could not source. If you manage a Business or Enterprise workspace, check your admin console. Kingy.ai says images with thinking are listed for Plus, Pro and Business, with Enterprise and Edu "coming later" per OpenAI.
• Real-world GIF quality. No benchmark covers it.

As of 13 November 2026, OpenAI's own announcement page was not accessible to me, so anything not tied to the developer documentation should be rechecked against it.

FAQ

What is ChatGPT Images 2.6?

It is OpenAI's image update released on 4 November 2026. It adds animated GIF creation, saved reference images and, in the API, two models (Flare and Sunburst), SVG output and markup-guided edits.

What is the difference between GPT-Image-2.6 Flare and Sunburst?

Flare is aimed at fast iteration, with OpenAI citing about six seconds at low quality. Sunburst is aimed at detailed work and consistency across edits at medium or high quality. Both use the same token rates.

Can ChatGPT make GIFs from my own photo?

Yes, according to launch coverage, you can give ChatGPT a prompt, photo or sketch and ask for a looping GIF. Rollout was still in progress as of 7 November, so it may not appear on every account yet.

Are Images 2.6 outputs labeled as AI-generated?

OpenAI says it uses C2PA metadata and SynthID watermarking, per its system card as summarized by Kingy.ai. I could not confirm this for GIF and SVG files, and metadata can be stripped on upload, so add your own visible disclosure where rules require it.

Is Images 2.6 more expensive than 2.5?

At high and max quality the price is unchanged. At low and medium it is higher, for example $0.02469 versus $0.00588 for a low-quality 1024 by 1024 image, according to OpenAI's calculator as reported by Kingy.ai.

Can I use saved reference images of people in ads?

I would only do so with written consent for commercial use. No source I opened describes OpenAI's consent safeguards for people, so the responsibility sits with you.

Does the SVG output replace a designer?

No. The API's SVG is static, generation-only and one file per request, and early testers called results uneven. Use it for quick icons and sketches, then have a designer finish anything that carries your brand.

Want help with this?

I'm Younus Fardeen, an organic growth, SEO and AEO strategist with 4+ years of marketing experience across edtech and startup teams. If you want help building a safe workflow for AI-generated visuals, with disclosure and rights rules that hold up, see my work and reach me through the contact form at younusfardeen.com.