ChatGPT Becomes a GIF Machine

Set Trending Topics as a preferred source on Google.

Sora is dead, long live the GIF: Eight months after OpenAI shut down its video app, the company is bringing motion back to ChatGPT another way. With its new image model, ChatGPT Images 2.6, users can turn a prompt or a photo into an animated GIF, edit it frame by frame and share it on a loop. The release also adds saved reference images and two new models for developers that OpenAI says are 50 percent faster.

How the GIFs Work

The idea is simple: Users describe a scene or upload an image, and ChatGPT turns it into a short animation. OpenAI demonstrates it with, among other things, a dancing cat. The key features:

• Creation: A text prompt, a photo or a rough sketch can serve as the starting point.
• Speed: The pace of the animation can be adjusted afterward.
• Frames: Anyone unhappy with part of the animation can rework individual frames.
• Sharing: The finished GIF plays on a loop and can be shared directly.

“From personal reactions to brand mascots and product visuals, GIFs give people and businesses a playful way to communicate,” OpenAI writes. The company has not said how long the GIFs can be, how many frames they contain or how many animations each subscription tier allows. Images 2.6 is rolling out to all ChatGPT and Codex users on the web, on desktop and on mobile.

Reference Images and Faster Models

Also new is a library for reference images. Users can keep reference images of themselves, their pets and their characters on hand and reuse them in new creations without uploading them again, OpenAI says. Characters, products and brand details are meant to look the same across multiple images. For companies that produce mascots or product shots in bulk, that may be the more important change.

Developers get two new models through the A.P.I. “Flare generates images in as little as six seconds,” according to OpenAI, while “Sunburst brings higher-quality generation and greater consistency across edits.” The release also adds image markup, which lets users mark the exact spots they want changed, and SVG export for websites, apps and print. OpenAI points to its own numbers to show how much demand there is: Users now create more than four billion images a week across ChatGPT and the A.P.I., up from three billion in September.

Motion Through the Back Door

Strategically, the move is notable. In March, OpenAI shut down Sora, its A.I. video app, which had also drawn criticism over deepfakes. That left the field to specialists like Higgsfield, which says it now makes $1 billion a year in revenue. GIFs are far less demanding to produce than videos, but they cover much of the moving imagery people share in chats and on social media.

That also puts OpenAI in the path of established GIF platforms like Giphy, which belongs to the stock image company Shutterstock, and Google’s Tenor. On those platforms, users pick from ready-made animations; in ChatGPT, the right reaction is made on demand.

The Competition

OpenAI is far from alone in the race for A.I. images. Google is pushing image editing in Gemini with its Nano Banana models, Meta has built its Muse Image model directly into Instagram, and Midjourney now offers short videos alongside images. In A.I. video, providers like Higgsfield, Runway and Google’s Veo lead the field. OpenAI also integrated the previous model, Images 2.5, into Adobe Firefly in September.

Animated Deepfakes at the Push of a Button

The new features raise familiar questions, however. Combining saved photos of real people with animation makes it even easier to show people in scenes that never happened. OpenAI’s announcements so far do not say how it safeguards GIFs of real people or whether the animations carry a watermark. With its Sora video app, the consumer advocacy group Public Citizen had demanded that OpenAI pull the app over the risk of deepfakes, and after protests from the actors’ union SAG-AFTRA, OpenAI tightened its rules for celebrities. Developers are not entirely enthusiastic either: In OpenAI’s developer forum, early users complain that low- and medium-quality images are billed at significantly higher rates through the A.P.I. than with the previous model.