OpenAI announced Images 2.6 on November 4, 2026, with global availability across ChatGPT and Codex subscription tiers on desktop, mobile, and web. API developers get two models: GPT-Image-2.6 Flare for faster generation and GPT-Image-2.6 Sunburst for higher-detail work.
The release adds native animated GIF generation, SVG output, markup-guided editing, and a complexity control separate from image quality. These additions expand the kinds of assets developers can request, while introducing different configuration, billing, and delivery requirements.
The practical distinction is between the broad product release and individual API capabilities. Images 2.6 has shipped, but some GIF-related API access may still be preview-gated. Developers should check their account permissions before promising animation features to customers.
Choosing Between Flare and Sunburst
OpenAI positions Flare around low latency and Sunburst around detail and editing consistency. Its published claims include image generation that is 50% faster than before and Flare generation in as little as six seconds at low quality.
Those are vendor claims, not independent measurements. Actual latency depends on the model, quality, complexity, and workload, so six seconds should not become an application’s guaranteed response time.
| Model | OpenAI’s positioning | Practical starting point |
|---|---|---|
| GPT-Image-2.6 Flare | Faster generation, including a claimed six-second minimum at low quality | Interactive drafts, quick variations, and workflows where waiting interrupts the user |
| GPT-Image-2.6 Sunburst | Higher-detail generation and more consistent edits at medium or high quality | Detailed assets and editing tasks where preserving visual information matters |
OpenAI’s system card also describes configurable reasoning effort, available through the ChatGPT model picker and API, as a way to influence generation and editing speed.
I’d start an interactive drafting workflow with Flare, then evaluate Sunburst for the final asset. That recommendation follows OpenAI’s model positioning; it is not a measured quality ranking. A production comparison should use the same prompts, dimensions, quality settings, and acceptance criteria for both models.
GIF Generation Has Its Own Delivery Rules
Images 2.6 adds animated GIF output, with API controls for the animation’s frame count, timing, and dimensions. The image-generation guide identifies three controls:
| Control | What it changes | Why it matters |
|---|---|---|
frame_count | Number of animation frames | Determines how many visual steps the animation contains |
frame_duration | How long each frame is displayed | Controls playback timing |
frame_size | Dimensions of each frame | Sets the animation’s output size |
Frame count and frame duration work together. Increasing the number of frames while keeping their display duration unchanged lengthens one playback cycle. Increasing each frame’s duration slows playback. Developers should follow the API’s accepted values and units when translating animation timing into request settings.
GIF output also has a distinct set of restrictions. OpenAI documents no support for:
- Streaming.
- Partial images.
- Custom output compression.
- Prompt rewriting.
These limits affect application design. A still-image interface that displays intermediate results cannot assume the same behavior for GIFs. An animation workflow needs a useful waiting state and a way to deliver the completed file.
I’d treat GIF generation as a background job for any customer-facing application where a long-running request could interrupt navigation. That is an implementation recommendation derived from the lack of streaming and partial output, not an OpenAI requirement.
The absence of custom output compression also removes one request-side lever for controlling delivery size. Developers with strict download budgets should inspect the resulting files instead of assuming generation settings will produce a predictable file weight.
Some GIF-related API access may be preview-gated despite the broader model launch. Verify that the intended account can make the required request before building the surrounding interface.
SVG Output Uses a Two-Stage Workflow
SVG support is more consequential than adding another file extension. OpenAI’s documented workflow first creates a reference image, then uses that image to produce SVG. The image quality setting controls the reference image. The resulting SVG also incurs charges for the model that authors the vector output. See the API guide.
That gives developers two stages to configure and evaluate:
- Reference generation: Is the composition, shape, and detail appropriate?
- SVG authoring: Does the vector file represent that reference in a useful form?
A detailed reference image does not, by itself, guarantee a simple or easily editable SVG. Vector structure matters when the asset will become an icon, logo component, or illustration that someone must adjust later.
The authoring-model setting is another setup detail worth checking. In the Hacker News discussion, commenter vunderba reported a 400 response when omitting svg.model, with the error:
“SVG generation requires 'svg.model'.”
That is an individual report, but it identifies a concrete integration issue: selecting an image model and requesting SVG may not be sufficient without configuring the authoring model.
For the returned file, OpenAI’s documented workflow is to decode data[0].b64_json from base64 and save the resulting UTF-8 XML bytes with an .svg extension. The response is carrying encoded SVG markup, so applications need to save it as the correct file type.
My view is that this workflow looks most useful where vector output has a clear downstream purpose, such as scalable interface artwork. For a photograph destined for a web page, the extra authoring step adds cost and processing without an obvious benefit.
Complexity and Quality Solve Different Problems
The new complexity parameter is separate from image quality. That distinction gives developers a way to control visual intricacy without treating quality as the only adjustment available in a request.
A clean application icon can need polished rendering while remaining visually simple. A dense illustration can require many elements even when it is still a draft. Those are different requirements, and combining them into one quality setting makes iteration harder.
The image-generation documentation describes complexity as its own control. For practical evaluation, hold the prompt and quality constant while changing complexity, then compare whether the output better fits the intended design.
I’d use that separation early in a workflow. If an asset is too busy, lowering quality is a poor substitute for asking for less visual complexity. The former changes the rendering target; the latter addresses the composition problem.
Markup-guided edits provide another form of control. OpenAI says Images 2.6 improves responses to sketches and image annotations, with photo markup supported in ChatGPT and the API. Its also describes improvements to detail preservation and editing consistency.
Frequently Asked Questions
3 questions
1How Do I Generate a GIF With Images 2.6?
Use the API’s GIF output workflow and configure frame_count, frame_duration, and frame_size. Check account access first, and plan for completed-file delivery because GIF output does not support streaming or partial images.
