What is new in ChatGPT Images 2.6?

Yesterday, Wednesday, November 4, 2026, OpenAI released Images 2.6, the latest version of the image generator built into ChatGPT, Codex (its coding assistant) and its API. The headline feature is motion: you can now ask ChatGPT for an animated GIF, the short looping image format used for reactions and memes. OpenAI also says images come out 50% faster, and that edits keep characters, products and brand details more consistent from one round to the next.

OpenAI says Images 2.6 is available globally, which includes Canada, on every ChatGPT subscription tier and on the web, desktop and mobile apps. The Verge reports that every ChatGPT and Codex user is getting it. The company also shared a usage figure: people now make about 4 billion images a week across ChatGPT, Codex and the API, up from 3 billion in September, according to The Verge.

What changed

Feature	Where	What it does
GIF creation	ChatGPT and API	Turns a prompt or a photo into a short animation
Faster generation	Everywhere	OpenAI says images are generated 50% faster than before
Better consistency	Everywhere	Characters, products and logos stay closer to the original across edits
Image markup	API (already in ChatGPT)	Lets you mark the exact spot on an image you want changed
SVG export	API	Outputs scalable vector graphics, which stay sharp at any size, for websites, apps and print
Complexity setting	API	Controls how much planning the model does before drawing

The Verge reports that in ChatGPT you can change a GIF's playback speed, look at and edit single frames, and then share it. Before this, ChatGPT could not make a GIF in one step. OpenAI pitches the feature at marketers animating a mascot, sellers turning a product photo into a loop, and animators sketching character movement before making a video. Even a rough drawing can serve as the starting point.

Flare and Sunburst: two models for developers

Images 2.6 keeps the two-model split OpenAI introduced with Images 2.5. The system card describes both as updated versions of the 2.5 models.

	GPT-Image-2.6 Flare	GPT-Image-2.6 Sunburst
Built for	Speed	Quality and precise edits
Best setting	Low quality, for drafts	Medium or high quality
OpenAI's claim	An image in as little as six seconds; cleaner layouts for posters and product shots	Higher image quality and more consistency across edits

Two settings now work together in the API. Quality controls how much fine detail the image gets. Complexity controls how much the model plans the scene before drawing it. For most edits, OpenAI recommends low complexity, so changes come back quickly, and medium or high complexity for crowded scenes with many elements or rules.

What it costs

On OpenAI's API pricing page, both models have the same price, which is also the price of the Images 2.5 models. Prices are in US dollars per million tokens, the units the model reads and produces, for text and for images.

Per million tokens (US$)	Input	Cached input	Output
Image tokens	$8.00	$2.00	$30.00
Text tokens (your prompt)	$5.00	$1.25	n/a

OpenAI does not list a flat price per image; it points developers to a calculator in its image generation guide. In ChatGPT, Images 2.6 is part of your existing plan.

Safety and labels

The system card reports how often each model produced a harmful image that the filters failed to block, in a test of 3,077 requests designed to break the rules. Sunburst improved slightly (1.72% against 1.92% for the 2.5 version). Flare got slightly worse (2.14% against 1.49%), though OpenAI says neither change is statistically significant. GIFs add a new difficulty: a series of frames can carry a meaning, or a suggestive movement, that no single frame shows. OpenAI says it added GIF-specific rules but that motion-related violations remain hard to catch.

OpenAI marks its images as AI-made in two ways. C2PA metadata is an industry-standard tag attached to the file. SynthID, Google DeepMind's invisible watermark, is an extra layer inside the image itself that does not depend on that metadata.

Here is a worked example of a GIF request in ChatGPT:

Prompt

Make a looping GIF of our café mascot, a round orange cat in a green apron, pouring coffee into a cup.
Six frames, side view, flat illustration style, cream background.
The last frame should flow back into the first. No text.

If the result is close but not right, ask for one change at a time: "slow it down", "make frame 4 show the steam rising", "keep everything else the same".

Try it: Draw a simple character on paper, take a photo, upload it to ChatGPT and ask for a three-second looping GIF of it waving. Then change only the speed and compare the two versions.

Should you use Flare or Sunburst?

Start with Flare. It is the faster model, and fast drafts let you try many ideas before you commit. Use it for brainstorming, social posts, quick layout tests and anything where you will look at several versions and keep one.

Switch to Sunburst when the image has to hold up: a product photo that must stay true to the real item, a character that must look the same across a series of edits, or a final asset for print. Since both cost the same per token, the trade-off is time, not price. A common approach is to explore with Flare at low quality, then redo the chosen idea with Sunburst at medium or high quality.