Animated GIFs, faster editing, and SVG export: what’s new in Images 2.6

OpenAI launches Images 2.6 with two models, Flare and Sunburst, a generation said to be 50% faster, animated GIFs, and new editing tools.

A mascot that comes to life, a poster whose composition can be corrected without rebuilding everything, or a visual exported as an SVG to be embedded in a website: image generation is no longer limited to producing an illustration from a description. With Images 2.6, OpenAI is extending its models to new formats and, above all, seeking to make successive edits easier, an essential step in professional creative workflows.

Announced on November 4, 2026, this version is available in ChatGPT, Codex, and the API. It introduces two models, GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst, which prioritize generation speed and detail quality, respectively. According to the information published by OpenAI, generation would now be 50% faster than on the previous generation, although no detailed comparative protocol was provided in the announcement.

This development comes as the company claims four billion images produced every week across ChatGPT, Codex, and its API, representing an average of more than 570 million creations per day. Two models to distinguish speed and precision

The new lineup distinguishes between uses requiring immediate results and those that prioritize a more detailed rendering.

Designed for rapid iterations, Flare can produce an image in six seconds under certain conditions, with the quality level set to “low.” The model also benefits from improvements in the arrangement of elements in posters, product visuals, and compositions with multiple graphic constraints.

Sunburst is more geared toward creations requiring precise detail rendering and better preservation of elements across successive edits. The company especially highlights the stability of characters, products, and visual identities across different versions of the same image.

For creative teams, this distinction can affect how work is organized. Initial graphic research can be done with a model that prioritizes speed, while the compositions selected may require greater precision during final edits.

However, these are not two models reserved for entirely different operations. Both support image creation and editing, as well as the new features introduced in version 2.6. Their behavior also depends on the parameters used. Animated GIF enters image generation

One of the main new features concerns the production of animations in GIF format. From a description, the system can generate several successive images and assemble them into a short animation, possibly designed to loop.

The examples presented include characters whose expressions change gradually, animated mascots, and illustrations intended for social media. A hand-drawn sketch can serve as a reference for creating an animated character.

In its official presentation, OpenAI notably shows two sequences created by OpenArt: a miniature office worker reacting to the fall and then the rise of a financial chart. Each animation is based on six successive images, arranged to preserve the same character, its environment, and a coherent progression of movements.

The format is also of interest to companies that want to animate graphic elements without undertaking a full video production. The startup Town, which develops a professional AI assistant, uses it, for example, to bring to life a character appearing in its interface.

The developer documentation specifies that the API can produce GIFs containing between two and 64 successive images, with a configurable duration between 20 and 2,000 milliseconds per image. The default setting corresponds to nine images at 200 milliseconds each.

This feature remains distinct from a video generation model. It produces an animation assembled from a sequence of images, without support for functions specific to audiovisual creation, such as audio generation. In the API, GIFs are currently created from the image generation endpoint, not from the one dedicated to edits. Localized edits and a new complexity setting

Beyond animation, the update introduces several features intended to better control changes made to an existing image.

The visual annotation system, already available in ChatGPT, is now accessible in the API. It makes it possible to directly designate a region to modify rather than describing only its location in a text prompt. This capability applies in particular to color changes, object corrections, and localized adjustments to a composition.

The editor also gains a parameter called “complexity,” independent of the generation quality level.

The setting determines the effort the model devotes to preparing the composition. A simple change, such as replacing a background or changing text, can be carried out with low complexity. A scene containing many objects, precise spatial relationships, or multiple graphic constraints may require a higher level.

The technical documentation indicates that high complexity can increase processing time and cause more unintended changes. The highest level is applied by default when this parameter is not specified.

This distinction between quality and complexity provides two control levers. The first affects the level of detail in the rendering, while the second influences the work needed to organize and preserve the scene’s elements. Choosing a higher complexity is therefore not systematically synonymous with a better result. SVG output for graphic uses

The creation of SVG files is another new feature intended for developers and design teams.

Unlike a classic raster image, an SVG document describes shapes and other graphic elements through a vector structure. It can be embedded in interfaces, websites, or print-oriented documents, with the ability to adapt to different sizes.

The proposed export relies on a multi-step process: a reference image is first produced, then a model tasked with transcribing it generates the corresponding SVG document. The result is therefore not necessarily an exact vector reproduction of every detail present in the initial image.

The API guide specifies that the resulting files contain static