gpt-image-2.6-sunburst Guide: Detail, Edits, SVG and GIF Output
gpt-image-2.6-sunburst is the detail-first model in OpenAI Images 2.6. Learn when it beats Flare, how complexity works, what SVG, GIF and transparency need, and how to migrate from 2.5.
gpt-image-2.6-sunburst is the model to reach for when an image has to be right on detail, layout, or edit accuracy, and it is not the model to run on every quick draft. When OpenAI launched Images 2.6 on November 4, 2026, it said Sunburst brings higher-quality generation and even greater consistency across edits when quality is set to medium or high. It is priced exactly like the faster gpt-image-2.6-flare, so there is only one reason to choose it: your acceptance bar is higher than what Flare reliably clears. This guide turns the official documentation into a decision method, covering how to set complexity, what SVG and GIF output can do, how to protect details across edit rounds, and how to move over from 2.5 Sunburst.
First decide whether the job really needs Sunburst
The Sunburst model page calls it image generation and editing for detailed visual work, and the prompting guide calls it the base model for workflows that need more detail or precise edits. OpenAI's selection rule is simple: for a new workflow, start with Sunburst if the image must meet demanding requirements for detail, layout, or edit accuracy, and start with Flare if response time matters most.
Jobs that sit closer to Sunburst's positioning:
Images with lots of small text, labels, charts, or relationships between objects, where one mistake makes the image unusable.
Edit chains where a person, a product, or a brand element has to stay stable across several rounds.
Formal ad creative, matched sets of product images, or layouts that go to a client.
Teams already on GPT Image 2.5 Sunburst, or teams whose GPT Image 2 workflow falls short on a complex task.
If your work is fast direction-finding with many candidates per round, read the gpt-image-2.6-flare guide first. You may not need to wait longer for Sunburst at all.
Same price, different use: compare cost per accepted image
OpenAI's pricing page lists gpt-image-2.6-sunburst at $5 per million text input tokens ($1.25 cached), $8 per million image input tokens ($2 cached), and $30 per million image output tokens. Those are the same rates as Flare and as both 2.5 models.
Matching rates do not mean matching cost per image. Images are billed by tokens, and model, size, and quality all change consumption. OpenAI's migration advice asks you to track failures, retries, and cost per accepted image. If Sunburst gets a difficult image through in one pass, it can be cheaper than repeated Flare attempts. If Flare already passes your checks, Sunburst's extra wait buys you nothing. Compare both on the same prompts, references, size, and quality, and let the numbers decide.
Keep expectations realistic too. The limitations section of the image generation guide says complex prompts can take up to two minutes, and that text rendering, character consistency, and precise composition can still go wrong. Sunburst raises the ceiling. It does not remove those problems.
Use complexity to control planning, and do not max it by default
The complexity parameter, new in 2.6, controls how much the model plans before it renders. The options are low, medium, and high. If you omit it, Sunburst uses high, and OpenAI says adjusting complexity has no separate fee. The separate quality setting controls rendering detail. The documentation stresses that higher complexity can increase latency and introduce unwanted changes, so you should compare results.
The prompting guide includes an experiment run on Sunburst. One prompt, a layered paper-cut illustration of a football team celebrating a championship, was rendered at 1024x1536, medium quality, PNG, changing only complexity. OpenAI's description of the outputs: at low, the trophy gets a crown. At medium, a laurel wreath decorates the cup and crests appear on the shirts, and both put a large banner in front of the team. At high, the "CHAMPIONS" sign moves into the stands, leaving the foreground for the players and the folds of their blue paper jerseys.
Three practical lessons follow:
Design decisions your prompt leaves open are where higher complexity gives the model room to plan.
When the prompt already fixes layout, text, and style, the guide says higher complexity may make little visible difference.
For a targeted edit, start with low and check that the surrounding details stay intact.
With Sunburst, raise complexity for open-ended scenes with many elements, lower it for tightly specified or local changes, and move one dial at a time.
Three outputs where precision matters: SVG, GIF, and transparency
The guide documents these deliverables, along with the limits that apply to them.
SVG. On the generation endpoint, set output_format to svg and n to 1 to get vector artwork with editable paths, shapes, and colors, which suits icons and illustrations. SVG does not support streaming, partial images, output_compression, or prompt_rewriter. The edit endpoint cannot output SVG, and an SVG is not a supported raster input for edits. OpenAI explains that the API first creates a reference image and uses it to produce the SVG, with quality controlling that reference image. The result is static vector content with no scripts, external resources, or embedded raster images. The official example runs one "Field & Flour" bakery logo prompt through both models.
GIF. Set output_format to gif and shape the animation with frame_count and frame_duration_ms. The documented defaults are 9 frames at 200 ms each, which makes a 1.8-second loop. OpenAI's example uses Sunburst for a monkey lifting a lantern at 512x512 and high quality. For a smooth loop, give the subject a starting pose to return to and say whether the camera should move. GIF is also generation-only, with no streaming, partial images, or custom compression. The guide and the API reference list different frame ranges, and the reference mentions access to a GIF preview, so confirm what your account can use before you design around it.
Transparent backgrounds. Ask for an isolated subject in the prompt and set background: "transparent" in the API, with PNG or WebP output. OpenAI warns that a drawn checkerboard is not transparency. Check the decoded alpha channel, especially at hair, glass, shadows, and object edges, and leave output_compression off for PNG.
Keeping details stable across edit rounds
Consistency is one of Sunburst's selling points, but OpenAI's own method is not "trust the model." It is to state constraints clearly and change one thing at a time. The patterns in the guide transfer directly:
Write edits as "change only X," then list what must stay: identity, geometry, layout, lighting, labels.
For clothing changes, name what cannot move: face, features, skin tone, body shape, pose, background, and camera angle.
Give every reference image a role: subject, style, clothing, or background.
Feed each output back as the next input, change one condition per round, and repeat the critical constraints. The official example generates a shampoo billboard, then adds only one sentence asking for a snowy winter evening.
Use markup to point at the spot to change. The overlay is a transparent PNG with the same pixel dimensions as the first input image. It cannot be combined with mask or prompt_rewriter, and you should omit input_fidelity.
One more official reminder deserves its own note: if a region must stay pixel-identical, composite the approved edit into the original. Prompting reduces drift. It does not guarantee zero drift.
Moving from 2.5 Sunburst
The six steps in the image prompting guide apply to Sunburst as they do to Flare. Save a baseline with hard edits, exact text, product geometry, and transparent assets. If you run 2.5 Sunburst, test 2.6 Sunburst first. Keep the first comparison identical in prompt, references, size, quality, and output format, with complexity at default. Check text, layout, identity, and the details that should survive an edit. Change either quality or complexity, never both at once. Then roll out by workflow in small shares while watching response time, failures, retries, and cost per accepted image.
Check the interface as well. OpenAI says GPT Image 2.6 does not support reasoning.summary, and you should omit input_fidelity on edits. On the model page I read, only image generation and image edit are marked supported while Responses and the other endpoints are not, so plan to call the Image API directly. The image generation guide also notes that you may need to complete API Organization Verification before using GPT Image models. If an older pipeline still runs gpt-image-1.5, OpenAI says it is scheduled to shut down on December 1, 2026, and gpt-image-1 was already shut down on October 23, 2026. For the previous generation of this decision, see the GPT Image 2.5 guide.
Common questions
What does gpt-image-2.6-sunburst do better than Flare?
OpenAI says that at medium or high quality, Sunburst gives higher-quality generation and greater consistency across edits. You still have to measure speed yourself. OpenAI's six-second figure applies to Flare at low quality, and no equivalent number is published for Sunburst.
How much does gpt-image-2.6-sunburst cost?
$5 per million text input tokens, $8 for image input, and $30 for image output, with cached input at $1.25 and $2. These rates match Flare.
When should I set complexity to high?
When the task has dense text, a complicated layout, or relationships between objects, and the prompt leaves design decisions open. OpenAI also warns that high can add latency and unwanted changes, so start with low for local edits.
Can Sunburst make a vector logo?
You can request SVG output on the generation endpoint with n set to 1. The result is static vector artwork. OpenAI tells you to review output before use, so finish brand-critical marks in a design tool.
