GPT Image 2.6: What Changes for Image Workflows

On November 4, 2026, OpenAI released GPT Image 2.6 Flare and Sunburst for image generation and editing, adding complexity control, SVG generation, and animated GIF generation. OpenAI positions Flare for fast iteration and Sunburst for detailed visual work. Their published token rates are identical—$5 per million text input tokens and $30 per million image output tokens—but your cost per finished image depends on actual token use and, for SVG, an additional authoring step. OpenAI’s dated changelog and image generation guide provide the release and workflow details.

What changed from GPT Image 2.5?

The November 4 release introduces two 2.6 model IDs and documents complexity control plus SVG and animated GIF generation. The same changelog says markup-guided editing was added for supported 2.5 and 2.6 models; it is not exclusive to 2.6. OpenAI’s September 8 entry had already introduced the 2.5 Sunburst and Flare models and their xhigh and max quality settings. Those quality settings should not be presented as a 2.6 invention. OpenAI’s changelog records both dates.

Question	Earlier 2.5 release	November 4 update
Model choice	Sunburst and Flare 2.5	Sunburst and Flare 2.6
Quality settings	xhigh and max already listed	Quality remains separate from complexity
Planning control	No 2.5 complexity control stated in the release entry	low, medium, high for both 2.6 models
SVG and GIF	Not part of the September 8 release entry	Generation routes documented for 2.6
Markup-guided editing	Not in the September 8 release entry	Added for supported 2.5 and 2.6 models

This is a documentation comparison, not a claim that every older workflow lacks a capability. The model tracker is useful for checking the two new model entries alongside other releases, while the maker’s documentation remains the source for API behavior.

Flare or Sunburst: which should you try first?

OpenAI describes Flare as an image model for fast iteration and Sunburst as one for detailed visual work. Those descriptions are positioning, not measured speed or quality scores. Start with Flare when you need to explore several concepts; include Sunburst in a controlled comparison when composition detail or edit precision determines whether an output is usable. Neither model page publishes a numerical speed advantage.

Published rate, per 1 million tokens	Flare	Sunburst
Text input	$5	$5
Cached text input	$1.25	$1.25
Image input	$8	$8
Cached image input	$2	$2
Image output	$30	$30

These are the rates shown on the two model pages. Equal rates do not prove equal bills: output tokens, reference-image input, quality choice, and the number of attempts can differ. Cached input rates on a pricing table are not a promise that a particular request will be cached. For a general explanation of why token classes matter, see our API price comparison and prompt caching guide.

For arithmetic only, suppose a hypothetical request uses 1,000 uncached text input tokens and 10,000 image output tokens, with no image input or other charges. At either model’s listed rates, the calculation is (1,000 × $5 + 10,000 × $30) / 1,000,000 = $0.305. Those token counts are assumed, not measured image usage or a price quote. A real request can consume different amounts, and an SVG request has another cost component. Use returned usage and your actual request settings to calculate spend. The site’s API cost calculator can help with token-rate arithmetic; it is not a calibrated quote for SVG or GIF production.

Complexity controls planning, not rendering quality

The new complexity request field tells the image model how much to plan before rendering. OpenAI’s guide lists low, medium, and high for both 2.6 models, with high as the default when omitted. The separate quality setting controls rendering quality. OpenAI suggests lower complexity for focused edits and higher complexity for dense layouts or relationships between objects. It also warns that higher complexity can increase latency and introduce unwanted changes.

That makes a small pilot more useful than automatically selecting the highest setting. For an edit that only changes a background, compare low with the default while keeping the source image and prompt fixed. For a dense poster layout, compare medium and high and inspect whether the model preserved the intended object relationships. These are suggested tests, not tests DailyTech conducted. OpenAI says complexity has no separate fee; the generation itself still uses billable tokens.

SVG and GIF require different output routes

The SVG instructions specify output_format: "svg" on POST /v1/images/generations. This produces static vector content intended for editable paths, shapes, and colors. Set n to 1; SVG output does not support streaming, partial images, output compression, or prompt rewriting. The image edit endpoint cannot output SVG, and an SVG is not accepted as a raster edit input. Decode the returned base64 as UTF-8 XML and save it as .svg.

SVG generation first creates a reference image, then an authoring model produces the vector document. OpenAI says you pay the image-model rate for the reference image and the authoring-model rate for its tokens. The final charge therefore needs both usage records. A static SVG is suitable when you need paths you can inspect and edit; it is not a shortcut for converting a current raster edit into an editable vector file.

For motion, the GIF instructions use output_format: "gif" on the same generation endpoint. The edit endpoint cannot output GIF. Specify the subject, motion, and loop in the prompt. The guide allows frame_count from 2 to 64 (default 9) and frame_duration_ms from 20 to 2,000 milliseconds (default 200, rounded to 10 ms). Frame count, frame dimensions, and requested animations must fit the request’s pixel budget. GIF output also lacks streaming, partial images, custom output compression, and prompt rewriting. Decode the returned base64 as .gif.

What to measure before switching

Use the same prompt, reference images, dimensions, and quality for a Flare–Sunburst comparison. Change only the model at first, then test complexity separately. Record wall-clock latency, returned token usage, total charge, and how many outputs meet your own acceptance criteria. For SVG, include the authoring cost and inspect the resulting paths; for GIF, inspect the loop and frame timing. This is a proposed evaluation method, not a benchmark result.

If your workflow depends on editing an existing image into SVG or GIF, the documented generation-only route is a design constraint to resolve before migrating. If it depends on a specific account limit or approval, check your own account: the public model pages say rate limits vary by usage tier. The release gives teams more format and planning options, but the decision still rests on outputs and costs measured in their own workload.

Sources

• OpenAI API changelog — November 4 release and September 8 comparison.
• GPT Image 2.6 Flare model page — positioning, token rates and endpoints.
• GPT Image 2.6 Sunburst model page — positioning, token rates and endpoints.
• OpenAI image generation guide — complexity, SVG, GIF and cost details.

All pages read on November 5, 2026.

Free toolWhat will 10,000 chats a month cost?Estimate my bill