Images 2.6 lands in ChatGPT with GIFs and editing

The news is not an independent ranking nor a promise that every image will come out ready for publication. OpenAI says the ChatGPT, Codex and API ecosystem generates 4 billion images per week, but that volume does not specifically measure Images 2.6.

In this article you will see:

• What changes in the Sunburst and Flare variants.
• How GIF creation starts from an idea or sketch.
• What the API controls allow you to adjust.
• How to test quality and recognize limits.

What Images 2.6 changes in image creation and editing

Images 2.6 brings generation and editing features together in a release focused on faster iteration and professionally finished images. OpenAI describes two modes of the GPT Image 2.6 family, called Sunburst and Flare, with different focuses.

Sunburst prioritizes higher quality images and greater consistency between edits when quality is set to medium or high. This consistency is useful when a character, product or brand needs to keep recognizable traits across multiple versions.

Flare focuses on layout composition for posters, product images and other professional visual materials. The distinction helps you choose the model according to the task, instead of asking a single adjustment to solve fidelity, speed and composition at the same time.

OpenAI presents Images 2.6 as faster to iterate than the previous generation, but the page does not turn that description into an independent scoreboard. To compare models, it is also worth looking at different tasks and methods, such as in the article about the MAI Image 2.7 in editing rankings.

The 4 billion images per week figure gives context to the scale of OpenAI's products. It includes images created in ChatGPT, Codex and API, so it should not be read as a quality, speed or specific adoption result for Images 2.6.

The GIFs of Images 2.6 can start from a sketch

The announcement also describes the creation of animated GIFs from instructions and visual references. The OpenAI page shows a simple hand drawn owl on lined paper as a starting point for an animated character.

OpenAI uses the example to show that a visual reference does not need to be polished final art. A short description can guide the transformation of a draft into an expressive character for videos, games or other projects.

This opens room for motion prototypes, campaign pieces and style explorations before investing time in finishing. Still, the generated GIF needs to be reviewed frame by frame when details, text, visual identity or animation continuity matter.

The announcement does not promise that any static image will become a consistent animation in every attempt. The result depends on the quality of the reference, the instruction and the review of the parts that need to stay the same.

The Images 2.6 API offers controls for distinct tasks

The update includes a complexity parameter to adjust how much planning the model dedicates to the request. OpenAI says smaller settings can handle quick edits, while requests with more elements can benefit from greater elaboration.

Quality controls the level of detail and visual fidelity, while complexity guides the content and the number of elements in the scene. Separating these choices helps test whether the task asks for just a localized correction or a more worked composition.

For teams integrating image generation into a product, these controls allow comparing instructions without changing the whole flow. The team can keep the same input image, vary one parameter at a time and record whether the edit preserved the items that should not change.

The OpenAI page itself notes that Images 2.6 is available in ChatGPT and the API. For the exact parameter names and accepted formats, see the official announcement page and the images documentation linked by it.

How to test Images 2.6 today without confusing quality with preference

Start with a test image you have permission to use. Write a single desired change and list separately the elements that must remain intact, such as colors, text, framing or product position.

Then compare one edit with another using the same reference. Change only the instruction or one control per round, and check small details, edges, letters and consistency between versions before selecting an output.

To experiment with GIFs, use a sketch or an image without personal data and describe the motion in concrete terms. Evaluate whether the character keeps shape and style across frames, not just whether the first frame looks correct.

The news also connects to other multimodal applications from OpenAI. See how the company describes the use of models in CAD, simulation and physical engineering, a different context from image creation for media.