GPT Image 2.6 Guide: Flare, Sunburst, GIFs, and API Features

OpenAI introduced Images 2.6 on November 4, 2026. The release brings faster image creation, more controlled editing, and GIF creation to ChatGPT, Codex, and the API. Developers can choose between two API models: GPT-Image-2.6 Flare for speed and GPT-Image-2.6 Sunburst for higher-quality results and more consistency across edits.

The headline claims need context. OpenAI says Images 2.6 generates images 50% faster than before, and that Flare can produce an image in as little as six seconds when quality is set to low. “As little as” describes a best-case result, not a guaranteed response time. Sunburst targets work where detail and continuity matter, especially at medium or high quality.

This guide explains what changed, how to choose between Flare and Sunburst, and how to use reference images, markup edits, the new complexity control, GIFs, and SVG output in a practical workflow.

GPT Image 2.6 at a glance

Flare vs. Sunburst: which GPT Image 2.6 model should you use?

Both models belong to the GPT Image 2.6 family, but they are designed for different priorities. Choose Flare when fast iteration matters: draft campaign directions, make routine product images, or explore a layout before spending time on a polished version. Its improved layout quality is useful for posters and product imagery, but inspect small text, spacing, and brand details before publishing.

Choose Sunburst when the image needs more detail or must remain recognizable through several edit rounds. OpenAI describes Sunburst as the higher-quality option and says it provides greater consistency across edits when quality is medium or high. That makes it a sensible starting point for a recurring character, a product scene that needs revisions, or a branded visual set.

These descriptions are starting points, not a guarantee that one model wins every prompt. For a fair comparison, keep the prompt, references, output size, and quality setting consistent. Review the final output for subject accuracy, composition, text, and details that should not have changed.

How fast is GPT Image 2.6 Flare?

OpenAI says Images 2.6 is 50% faster than before. For the API, the specific minimum called out in the launch is as little as six seconds with Flare at low quality. It is best read as a lower-end example, not as an average or a service-level guarantee. Higher quality, larger images, editing inputs, and complex scenes can change latency.

If speed matters to your application, measure your own requests rather than relying on a single headline number. Record the model, prompt, image dimensions, quality, complexity, whether source images were attached, and response time. Compare representative requests and look at both median and slower-tail results. This shows whether Flare improves the workflow you actually run.

Reference images and consistent edits

Reference images help carry visual information into a new generation or edit. They can show the product to preserve, the character to reuse, a desired material, or a composition to follow. State the role of each reference in the prompt instead of assuming the model will infer it. For example: “Use the first image for the bottle shape, the second for the label colors, and create a new clean studio background.”

When a series of images needs continuity, identify the fixed details explicitly: silhouette, markings, color palette, camera angle, and any text or product geometry that must remain. Then list what is allowed to change. Sunburst is the natural option to evaluate when edits need more consistency, but check the output after each round. “Greater consistency” does not mean every feature will be preserved perfectly.

Image3 has more background on the GPT Image 2 workflow, a step-by-step guide to using GPT Image 2, and a hands-on AI image generator for creating and refining images.

Markup editing and the new complexity control

Markup editing is useful when a text instruction alone is too broad. In the Image API edit workflow, provide the source image, a transparent PNG annotation layer, and a prompt describing the intended change. The markup indicates where the model should focus; the prompt explains what to do there. For example, circle a vase and ask the model to recolor it while keeping the rest of the room unchanged. The annotation is a guide, so review the finished image to make sure it has not left marks or changed unrelated details.

GPT Image 2.6 also adds a complexity request parameter. It controls how much planning the model does before rendering; quality controls visual detail and fidelity. They solve different problems. OpenAI recommends low complexity for most edits and medium or high complexity for scenes with many elements or constraints. The API guide says complexity supports low, medium, and high, defaults to high when omitted, and does not add a separate fee. Set it intentionally: lower complexity can make focused edits faster, while a dense composition may need more planning.

For edits to a particular area of an existing picture, Image3 also offers an AI inpainting workflow.

GIF creation and SVG export

Images 2.6 adds GIF creation for ideas that benefit from motion: a looping mascot reaction, a short product animation, or a storyboard that helps a team explore movement before producing a video. Keep the sequence simple and describe what should remain fixed from frame to frame. The launch announcement includes GIF creation as part of the Images 2.6 experience; confirm the exact output format supported by the ChatGPT, Codex, or API workflow you plan to use before building a production pipeline around it.

For scalable artwork, the API guide documents SVG output on the image-generation endpoint. SVG is useful for editable icons, illustrations, and simple logos that need crisp scaling. The model creates a reference image and uses it to produce static vector content. The documented SVG workflow requires n to be 1; it does not support streaming, partial images, output compression, or the prompt rewriter, and the image-edit endpoint does not return SVG. Treat the vector as a generated starting point and inspect its paths, shapes, colors, and text before using it in a production design system.

A practical GPT Image 2.6 API workflow

Start with the fastest setting that can answer your question, then increase fidelity only when the task needs it. A focused Flare draft might look like this in Python:

result = client.images.generate(
    model="gpt-image-2.6-flare",
    prompt="Create a clean studio product image of a cobalt ceramic vase on a pale stone shelf.",
    quality="low",
    complexity="low",
)

For a final visual where fine detail matters, test gpt-image-2.6-sunburst with a medium or high quality setting. For a localized edit, use the edit endpoint with your source image and markup layer. For a complex poster with several objects, try medium or high complexity, then check whether the added planning actually improves layout and text placement.

Use this checklist before choosing defaults for an application:

1. Separate draft and final requests. Use a low-quality Flare draft for quick direction checks; use the model and quality setting that meet the delivery brief for the final image.
2. Make references explicit. Tell the model what to preserve from each input image and what to change.
3. Set complexity deliberately. Try low for a narrow edit; compare medium or high for a dense scene with multiple constraints.
4. Use markup for location. A visual annotation can remove ambiguity about the target region. Keep the instruction in the prompt specific.
5. Measure latency and cost. Image output is billed by token usage, not by one universal flat price per image. The published output-token rate is $30 per million image-output tokens; total cost also depends on output size and quality, plus any text or image inputs. Use response usage data and the pricing calculator for your actual request mix.
6. Keep a regression set. Re-run representative prompts after changing models or settings. Compare output fidelity, edit preservation, layout, time, and cost.

Frequently asked questions

Is GPT Image 2.6 released?

Yes. OpenAI announced Images 2.6 on November 4, 2026. It is available in ChatGPT and Codex across subscription tiers, and the API offers GPT-Image-2.6 Flare and Sunburst.

What is the difference between GPT Image 2.6 Flare and Sunburst?

Flare is the speed-focused API option and can generate an image in as little as six seconds at low quality. Sunburst is aimed at higher-quality generation and greater consistency across edits at medium or high quality.

Does GPT Image 2.6 generate images 50% faster?

OpenAI reports a 50% speed improvement for Images 2.6 compared with before. This is a launch-level claim; the time for an individual request depends on its model, settings, input images, and scene complexity.

What does the complexity parameter do?

It adjusts how much planning the model applies before rendering. It is separate from quality, which controls visual detail and fidelity. The API supports low, medium, and high complexity; the documented default is high.

Can GPT Image 2.6 use markup, references, GIFs, and SVG?

The API guide documents image references and markup-guided edits for supported GPT Image models, and SVG output for image generation. GIF creation is part of the Images 2.6 launch. Check the selected surface and endpoint before assuming all formats share the same output behavior; SVG, for example, is not returned by the image-edit endpoint.

Feature details checked against OpenAI’s public launch and API materials on November 5, 2026. The cover and five inline images are original editorial illustrations; they are not official OpenAI screenshots or model outputs.