On 4 November 2026 OpenAI added gpt-image-2.6-sunburst and gpt-image-2.6-flare to the Image API. Generation uses POST /v1/images/generations; edits use POST /v1/images/edits. GPT Image 2.6 adds adjustable complexity, output_format values svg and gif, and markup-guided edits shared with GPT Image 2.5. GPT image models always return data[].b64_json—do not send legacy response_format: url.
Dated 5 November 2026 from the 4 November changelog and the image generation guides. Two days ago we covered Anthropic message threads—incremental Messages JSON. Here the change is new output formats and edit fields on the image endpoints. If you already validate 2.5 bodies, swap the model id and add separate schemas for GIF, SVG, and markup edits.
Sunburst vs Flare and shared fields
gpt-image-2.6-flare favors latency; gpt-image-2.6-sunburst favors detail and edit accuracy. Both accept complexity (low, medium, high; docs default high when omitted), quality through max, transparent background, and raster png, jpeg, or webp.
{
"model": "gpt-image-2.6-flare",
"prompt": "Flat vector icon of a lantern, centered",
"size": "1024x1024",
"quality": "high",
"complexity": "medium",
"output_format": "png",
"n": 1
}
Set the model only on the Image API. In Responses, put the 2.6 id on the image tool’s model field. Chat Completions does not host these snapshots.
GIF generation JSON
Set output_format: "gif" only on /v1/images/generations—the edit endpoint does not emit GIFs. Control animation with frame_count (2–64, default 9) and frame_duration_ms (20–2000, default 200, rounded to 10 ms in the file). GIF output disallows streaming, partial images, output_compression, and prompt_rewriter.
{
"model": "gpt-image-2.6-sunburst",
"prompt": "Paper lantern swaying in a loop, same start and end pose",
"output_format": "gif",
"frame_count": 9,
"frame_duration_ms": 200,
"size": "1024x1024",
"n": 1
}
size is each finished frame. Frame count × frame pixels × n must fit the model pixel budget—reduce frame_count or dimensions, or use auto. Your org needs GIF preview access for the chosen model.
SVG generation JSON
On generations, set output_format: "svg". Require n: 1. SVG disallows streaming, partial output, output_compression, and prompt_rewriter, and cannot be used as a raster edit input. Decode data[0].b64_json to UTF-8 XML static vectors without scripts or external assets.
{
"model": "gpt-image-2.6-sunburst",
"prompt": "Minimal bakery logo, wheat and rolling pin, single color",
"output_format": "svg",
"quality": "high",
"complexity": "high",
"size": "1024x1024"
}
Responses may include separate svg_usage token fields from the SVG author model—keep them out of your raster usage schema.
Markup-guided edit JSON
GPT Image 2.5 and 2.6 models accept markup on /v1/images/edits: a transparent PNG overlay matching the first input image dimensions. JSON edits pass the base photo in images; markup is an object with exactly one of image_url or file_id (HTTPS or base64 data URL).
{
"model": "gpt-image-2.6-flare",
"prompt": "Replace the circled vase with blue ceramic, keep kitchen lighting",
"images": [
{"image_url": "https://cdn.example/kitchen.webp"}
],
"markup": {
"image_url": "data:image/png;base64,iVBORw0KGgo..."
},
"size": "1536x1024"
}
Never combine markup with mask or prompt_rewriter. Markup applies only to the first input; extra images entries are references. Multipart uploads use a file field for markup—keep that out of your JSON schema.
Response JSON and usage
Success payloads include created and a data array whose items carry b64_json for PNG, GIF, or SVG alike. GPT image models do not return revised_prompt or hosted url fields.
User-correctable failures may set error.type to image_generation_user_error with stable error.code values such as moderation_blocked and optional moderation_details. Do not blindly retry—change the prompt or inputs first.
Validating request bodies locally
Maintain at least four schemas: generations/raster, generations/gif, generations/svg, and edits/markup-json. Use JSON validation to catch frame_count on non-GIF bodies or markup objects that include both image_url and file_id.
FAQ
Use 2.6 as the top-level Responses model?
No. 2.6 runs on Image API routes only; Responses references it on the image tool.
GIF or SVG from the edit endpoint?
No. Only generations emit GIF or SVG; edits return raster.
Still send response_format url?
Not for GPT image models. Expect b64_json only.
Markup plus mask in one edit?
No. Pick markup or mask, not both.
frame_count on edit JSON?
Invalid. GIF frame fields belong on generations with output_format gif.
Same JSON as ChatGPT one-step GIF?
The product UI wraps prompts; the API still needs explicit gif fields and org access.
Summary
GPT-Image-2.6 separates model choice and complexity from output_format, frame controls, and markup—validate generations and edits with different JSON schemas.