GPT-Image-2.6 Rolls Out Everywhere With Native GIF, SVG and a Separate Complexity Dial

OpenAI's image model has a new default. GPT-Image-2.6 is now fully live in ChatGPT, Codex and the API, open to free and paid users alike, with total generation time cut by roughly 50 percent and six-second renders in the fastest setting, according to Chinese tech outlet QbitAI.

The most visible additions are motion and vectors. The model generates GIFs natively and exports SVG directly, removing the third-party bitmap-to-vector step that front-end and UI work usually requires.

The more consequential change sits in the API's control surface. OpenAI split the compute spent planning a scene from the compute spent rendering it: a new complexity parameter governs how much work the model does on composition before it draws, while quality now covers only visual detail and fidelity. Previously, changing a background colour forced a full re-plan of the layout.

Two API variants come with the split. Flare targets speed and layout for posters and e-commerce product shots, reaching the six-second mark at the lowest quality tier, while Sunburst covers medium and high quality and focuses on keeping characters and products consistent across multi-turn edits.

GIFs are produced without invoking a video model. The published OpenArt example prompt asks for a single sprite sheet holding six sequential frames in a 3x2 grid, which is then sliced and assembled. That sidesteps the inference latency of video diffusion and keeps faces, fabric folds and background objects fixed from frame to frame.

Adoption is already starting. Town, a startup building an AI office assistant, uses the API to generate real-time micro-expression GIFs — typing, feedback cues — for virtual assistants embedded in work software.

Two engineering additions round it out. Native SVG export lets developers working in Codex produce infinitely scalable icons, components and illustrations straight into a codebase, and Image markup, previously limited to the ChatGPT web app, is now exposed to the API so teams can target regions by coordinate or mask and run batch edits at Sunburst's consistency.

Read together, OpenAI is selling more than faster rendering: it is selling a set of image-production parameters a workflow can actually tune. For developers, the cost granularity just got finer; for competitors, the unit of comparison shifts from the price of an image to the compute spent on a single revision.

The report closes with an unconfirmed aside: an observer reportedly spotted a model entry labelled GPT-6.2 Astra in Google Cloud. It has no stated connection to the image release, and OpenAI has said nothing publicly about whether a new model is near.