What OpenAI shipped on 4 November

OpenAI's API changelog lists two new image models dated 4 November 2026, GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare. The new features are a complexity setting, SVG output, animated GIF output and markup-guided editing. According to the image generation guide, Sunburst is for detailed work where editing precision counts and Flare is for fast iteration. The guide tells developers starting new integrations to use one of the 2.6 models.

MarkTechPost reports OpenAI's claim that Flare can return an image in as little as six seconds at low quality, and that Images 2.6 is 50% faster than the previous release. OpenAI's launch post returned an error when we tried to open it, so we couldn't check those figures at the source. Treat them as vendor claims until your own timings agree.

Endpoints, versions and the complexity setting

Both models take text and image input and return images. Both run only on v1/images/generations and v1/images/edits. The Flare model page lists Batch, Responses and Chat Completions as unsupported, and the Sunburst page says the same.

The missing Batch support affects bulk jobs. A DEV Community analysis notes that GPT-Image-2.5 Flare still supports Batch at $15 per million image output tokens. Version 2.6 charges $30 at standard rates, so a large overnight catalogue job may still be cheaper on 2.5. The same article notes that each 2.6 model lists only one snapshot, its own alias. You can't pin a dated version yet.

complexity takes low, medium or high, and high is the default when the field is left out. The setting controls how much the model plans before rendering. quality is a separate setting that controls rendering detail. The guide suggests low for small, targeted edits such as replacing a background or the text on a sign. It suggests high for dense text, complicated layouts and scenes where objects relate to each other. The guide also warns that higher complexity can add latency and make changes you didn't ask for. Complexity has no separate fee, so the cost of testing all three levels on your own prompts is ordinary token use.

SVG output for icons and diagrams

To get vector output, send output_format: "svg" to the generations endpoint with n set to 1. Per the guide, the request runs in two stages. The image model renders a reference image, and an authoring model then writes the SVG. The API reference adds that complexity only affects the reference image and has no effect on how the SVG author reasons. Billing follows the same split: you pay the image model's rate for the reference image plus the authoring model's rate for the tokens it uses to write the SVG.

The SVG arrives base64-encoded in data[0].b64_json. Decode it and save the UTF-8 XML with a .svg extension. The guide describes the output as static vector content with no scripts, external resources or embedded raster images. That makes it easier to sanitise and ship than arbitrary SVG from the web. SVG requests don't support streaming, partial images or output_compression. The edits endpoint can't return SVG, so you can't send in an existing raster logo and get a vector version back. The model draws each SVG from scratch based on your prompt.

Animated GIFs

output_format: "gif" works on both the generations and edits endpoints. The guide suggests the edits endpoint when you want still reference images to guide the animation. The GIF settings are:

  • frame_count: 2 to 16 frames, default 9
  • frame_duration_ms: 20 to 2000, rounded to the nearest 10 ms, default 200
  • size: auto or WIDTHxHEIGHT in multiples of 16, applied to each frame

The API reference says GIF output requires access to the GIF preview for the chosen model. GIF requests also don't support streaming, partial images, masks or custom compression. Check that your organisation has preview access before you promise a client an animated feature. The DEV Community article saw an older guide that allowed up to 64 frames. Both pages we read now give 16 as the limit.

Markup-guided edits

Markup is a transparent PNG laid over the original image, with circles, arrows or sketches showing where the change should go. You describe the change in the prompt. This shortened example is from the guide:

result = client.images.edit(
    model="gpt-image-2.6-flare",
    image=("kitchen.webp", image, "image/webp"),
    markup=(overlay.name, markup, "image/png"),
    quality="medium",
    prompt="...",
)

The documented steps:

  1. Send the original image as the first input image.
  2. Draw the annotations on a transparent PNG with the same width and height as the original after orientation. Keep it as a separate file and don't merge it into the original.
  3. Send that PNG as markup. Multipart requests take a file upload. JSON requests take an object with exactly one of image_url or file_id.
  4. In the prompt, describe the change and what should stay untouched, then review the result.

Markup only applies to the first input image, and any further images serve as references. The original and the overlay are each limited to 16,777,216 pixels, and the overlay file to 50 MB. You can't use markup and mask in the same request.

Costs and limits

Both models have the same per-million-token prices: text input $5 ($1.25 cached), image input $8 ($2 cached) and image output $30, all in US dollars (Flare page). OpenAI hasn't published a per-image price for 2.6. Each partial image adds 100 image output tokens. The guide says cached input rates only apply through the Responses API image tool. Because the 2.6 model pages list Responses as unsupported, we wouldn't count on cached rates until OpenAI clarifies.

Per-image cost can still rise even though the token rates match 2.5. The DEV Community article cites one forum user who counted 1,136 output tokens for a medium image on 2.6 against 344 on 2.5. That's a single unofficial report, so log the usage field on your own requests before you set prices.

The guide also says complex prompts can take up to two minutes. Text rendering can still be imperfect, and recurring characters or brand elements may drift between generations. Content filtered by moderation returns moderation_blocked. Don't retry those requests until the prompt or input image has changed. Rate limits depend on your usage tier, and the model pages don't give numbers.

What builders here could make

The models are OpenAI-hosted only, according to MarkTechPost. None of the pages we read mention options for processing in a particular region. Teams with data residency requirements should check that before sending client images.

The SVG path suits product teams that need consistent icon sets or simple diagrams to commit straight into a component library. Markup edits suit tools where non-technical staff circle what should change, such as retail product photos or property listings. A lightweight review step catches the drift the guide warns about. GIF output could feed social posts for small Auckland businesses, as long as the preview access is in place.

If you build one of these, you can share it with other members on our projects page.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member