OpenAI released Images 2.6 on 4 November. The update reaches ChatGPT, Codex, and the API in one move. Three changes matter to designers more than the speed claim in the headline: native GIF output, SVG export, and a split between two models that trades speed against consistency.
The ChatGPT side: GIFs in one step
You can ask ChatGPT for a looping GIF from a text description, or upload a still and ask it to animate. After generation you change playback speed, edit by describing changes in the conversation, and save or share the file. One limit: ChatGPT cannot edit an animated file you upload. The feature rolls out gradually across every ChatGPT tier on web, desktop, and mobile.
OpenAI also mentions saved reference images of people, pets, and characters that carry across creations. If you run a recurring mascot or a founder headshot through social assets, that reference system saves you re-uploading and re-describing.
The API side: Flare drafts, Sunburst holds the line
Developers get two new models instead of one. GPT-Image-2.6 Flare targets fast iteration. At the low quality setting it returns an image in about six seconds, and OpenAI says it handles poster and product-shot layouts better. GPT-Image-2.6 Sunburst targets detailed work. At medium or high quality it aims to keep characters, products, and brand elements consistent from one edit to the next.
Read that as a workflow: draft with Flare, finish with Sunburst. Six seconds per draft changes how many directions you try before you commit. Consistency across edits is the feature every brand manager has asked image models for since 2023.
The parameters worth knowing
- GIF output through the Image API: set the output format to GIF, choose 2 to 64 frames (nine by default), and set per-frame duration between 20 and 2,000 milliseconds. The edit endpoint does not accept GIF output.
- Image markup: mark the exact region to change before the call. Targeted edits without a mask file.
- SVG export: the API returns vector files for websites, apps, and print. The edit endpoint does not support it yet.
- Complexity parameter: sets how much planning the model does before drawing, separate from the quality setting. OpenAI recommends low complexity for quick edits and higher values for crowded scenes.
- Speed: OpenAI claims images generate about 50% faster than the previous version. No independent benchmarks accompany the release.
A view from a local pipeline
This blog generates every header image on a local GPU with Qwen Image 2.1 through ComfyUI. That stack gives full control, no per-image cost, and no data leaving the machine. It also gives you no vector output, and character consistency across edits is something you build yourself with reference images, seeds, and LoRAs.
Images 2.6 closes two of those gaps on the API side. SVG output means a generated mark scales to print without a trace-and-clean-up pass. Sunburst's consistency promise means a product shot edited five times still shows the same product. If your work ships logos, icons, or print collateral, OpenAI just removed steps from that pipeline that local models still ask you to do by hand.
Scale is the other signal. OpenAI reports people now create 4 billion images per week across its surfaces, up from the 3 billion it cited in September. Figma Weave and OpenArt already ship the model in their tools. Our Midjourney vs Flux comparison covers the prompt-control side of this market; Images 2.6 moves the API side forward on the two features designers kept requesting.
ChatGPT's saved references hold faces and characters, but brand specs still need one source of truth. We keep our token values and reference rules in Memkist so every image tool reads the same values, and Asa picks which model answers each task. Decide where your brand source lives before you add another model to the rotation.
Test before you switch
Run three checks with your own briefs. Draft ten concepts on Flare and time the loop against your current tool. Edit one product shot five times on Sunburst and inspect what drifts. Export an SVG and open it in Illustrator or Figma to see how clean the paths actually are. OpenAI's numbers come from OpenAI. Your brand is the benchmark.

