ChatGPT Images 2.6 adds animated GIFs, saved references and two faster API models

OpenAI announced Images 2.6 on November 4, an update to the image generator built into ChatGPT. Its most visible addition is animated GIF creation. According to The Decoder and TestingCatalog, the update is rolling out to all ChatGPT and Codex users, on every subscription tier, across the web, desktop and mobile apps.

GIFs and saved references in ChatGPT

Users can start from a written prompt or an uploaded photo and ask for a short animated loop. Once a GIF is generated, they can change the playback speed or edit single frames before sharing it. The Decoder describes OpenAI's demonstration, a dancing cat rendered as a series of slightly different images played in sequence, and argues that the feature depends on the model keeping a character recognizable from one frame to the next, something earlier models struggled with. TestingCatalog lists mascots, product shots, sketches and storyboards among the intended uses.

A second feature lets people save reference images of themselves, their pets or recurring characters once and call them up in later creations without uploading them again. OpenAI also says usage keeps growing: more than four billion images a week are now generated across ChatGPT and the API, according to The Decoder, about a billion more than in September.

Flare and Sunburst for developers

On the developer side, OpenAI's forum announcement names two models, gpt-image-2.6-flare and gpt-image-2.6-sunburst, and says image generation in the API is 50% faster. Flare is the speed-oriented option and, OpenAI says, can return an image in as little as six seconds; TestingCatalog notes that this figure applies at low quality and that Flare is aimed at layouts such as posters and product imagery. Sunburst is meant for higher-quality output and more consistent results across successive edits. The API also gains image markup, already present in ChatGPT, which lets developers mark the exact areas to change, as well as SVG export and a new parameter that controls how much planning the model does before generating.

Open questions

The announcement links to OpenAI's pricing page but does not state prices in the post itself. In replies on the forum, a developer reported that low and medium quality settings appeared to produce far more billed output tokens than the previous version, though these observations are not official figures. The Decoder points out that the Flare and Sunburst names first appeared with Images 2.5, and the forum thread notes that parts of the documentation still referred to 2.5 model names at launch.