ChatGPT Images 2.6 adds GIFs and saved reference photos. Who gets it, and what it costs

What OpenAI released

On November 4, 2026, OpenAI introduced Images 2.6, saying it’s “available globally today in ChatGPT and Codex across all subscription tiers on desktop, mobile, and web.” It generates images “50% faster than before,” and OpenAI says it holds details steadier across edits. The company also says people now make 4 billion images a week across its products.

ChatGPT gets two new features:

• GIFs. The release notes say GIF creation “is rolling out,” so it may not be on your account yet, even though the model is.
• Saved reference images. You save a photo once and call it by name in later chats, instead of uploading it again.

Developers get two models, gpt-image-2.6-flare and gpt-image-2.6-sunburst, plus GIF and SVG output, a complexity setting and markup-guided edits.

How to use it in ChatGPT

OpenAI’s help page gives the steps.

1. Make a GIF. Ask for a looping GIF and describe what should move, or upload a still image and ask ChatGPT to animate it.
2. Change it. Describe the edit in the conversation. Open the GIF to adjust its playback speed, save it or share it. ChatGPT “can’t edit an animated GIF you upload.”
3. Save a reference. Upload a photo, ask ChatGPT to save it as a reference, and choose a name. ChatGPT may also offer when you describe a photo, such as a picture of your dog.
4. Reuse it. In a later chat, type @ and the name: “Make a watercolor portrait of @Pepper.”

The Verge reports that you can also edit individual frames, and that an example from OpenAI shows a frame-by-frame view. The help page doesn’t describe frame editing yet, so you may not see it.

Thinking is a separate gate. Images with thinking is on Plus, Pro and Business and “coming to Enterprise and Edu soon,” and the pricing page lists it as unavailable on Free and Go.

What it costs in the API

The token prices didn’t change from 2.5. These are OpenAI’s standard rates per million tokens, as listed on November 5.

Model	Image in	Image out	Cached image in	Text in	Best for
gpt-image-2.6-flare	$8	$30	$2	$5	Fast iteration, posters, product imagery
gpt-image-2.6-sunburst	$8	$30	$2	$5	Detail and consistency across edits
gpt-image-2.5-flare and -sunburst	$8	$30	$2	$5	Older versions at the same rate
gpt-image-2	$8	$30	$2	$5	Older model at the same rate

The same rate doesn’t mean the same bill. OpenAI’s guide warns that “equal token rates don’t mean equal cost per image,” because token use differs by model and quality, and it has a calculator for estimating an image.

The pricing page’s Batch table, at half these rates, lists gpt-image-2 and both 2.5 models but neither 2.6 model. If you depend on Batch, check before you switch.

What’s new for developers

These work through the Image API, and the guide has the limits.

• GIFs. Set output_format: "gif" on POST /v1/images/generations. frame_count takes 2–64 frames and defaults to 9, and frame_duration_ms takes 20–2000 and defaults to 200, so the default loop runs 1.8 seconds. The edit endpoint can’t output GIFs.
• SVG. Set output_format: "svg", with n at 1. The result is static vector content, and the edit endpoint doesn’t support it.
• Complexity. low, medium or high sets how much the model plans before rendering, at no separate fee. If you omit it, you get high, which the guide says “can increase latency and introduce unwanted changes.” OpenAI recommends low for most edits, so set it deliberately.
• Markup. Send an annotation layer as markup on POST /v1/images/edits to point at what should change. It also works on supported 2.5 models, and can’t be combined with mask.

There’s a date attached if you’re on an older model. OpenAI’s deprecations page removes gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest from the API on December 1, 2026. It names the 2.5 models as replacements and hasn’t been updated for 2.6. We’d test 2.6 first, since it costs the same per token and the page gives no shutdown date for either.

What about Codex

OpenAI’s documents disagree here, so don’t assume.

Its announcement puts Images 2.6 in Codex, and TestingCatalog quotes OpenAI Developers saying the same. Its Codex image page still said on November 5 that “built-in image generation uses gpt-image-2.” That page also has the part that matters for your plan: images count toward your general Codex limits and use them “3–5x faster on average” than similar turns. For larger batches, it says to set OPENAI_API_KEY so API pricing applies.

GIFs look further off in the CLI. A pull request merged on November 3 added GIF output to Codex’s image tool, with 2–16 frames, but the feature flag was still marked under development and off by default on November 5. That’s only the open-source code. We can’t see how the desktop app behaves.

Safety, and what OpenAI hasn’t said

OpenAI’s system card says GIFs bring risks a single frame doesn’t show, such as “sexualized movement,” so some categories are blocked more broadly than in still images. On a deliberately hard test set, its GIF safeguards caught 77.7% of violating GIFs overall, and it says “motion-related violations remain challenging to detect.” That isn’t a rate for real use. Images also carry C2PA metadata and an invisible SynthID watermark.

We couldn’t find answers to these on OpenAI’s pages on November 5:

• How many GIFs each plan gets, or whether a GIF counts as one image against your limit.
• How long a ChatGPT GIF can be. The 2–64 frame range is the API’s.
• What a GIF costs in the API. The cost calculator covers still images.
• How to delete a saved reference. The help page says how to save and use one, not how to remove it.
• When GIF creation will finish rolling out.