What GPT Image 2.6 Sunburst and Flare actually ship
Two variants. According to AI Dev Weekly, the image generation guide positions Sunburst for detailed visual work and Flare for faster iteration. Both support SVG and animated GIF output through the Image API. So the choice is detail or speed, and you make it per call.
Scott Harvanek's 72-hour brief dates the release to November 4 and quotes OpenAI's own summary: faster generation, GIF creation, more precise editing and new API controls. Note that "faster" is OpenAI's claim. The sources we read give no latency figures, so we have nothing to check it against. We'd treat speed as unproven until we've timed our own prompts.
The complexity setting is a second dial, and it defaults to high
The new complexity setting accepts low, medium or high, and the default is high. It controls planning effort separately from the rendering quality setting. Two dials is a sensible design. In theory, a simple edit shouldn't pay for scene planning it doesn't need.
The default is the part to watch. If you port an existing call and change only the model name, you get the highest planning setting whether or not the job needs it. We'd set complexity explicitly in every call and start low for small edits.
What we can't tell you is how much complexity moves price or latency. The sources we have don't say. Check OpenAI's pricing page, then measure on your own prompts. Compare cost per accepted asset, not cost per request. A fast model that needs three retries isn't fast.
SVG and GIF output are useful, and still untrusted files
Native vector and animated output means some assets no longer need a separate conversion step. That's real. Icons, spot illustrations and simple animated banners are the obvious candidates.
An SVG from a model is still XML from a model. AI Dev Weekly advises checking that it renders in your target browser and sanitizing it before serving it as active content. We agree, and we'd go further: treat it like a user upload, because SVG can carry more than shapes. For a GIF, check file size, legibility and playback in the places customers will actually see it.
There are open questions. The sources don't say whether SVG output can follow existing artwork, or how closely it can match a brand style. Test that before you promise anyone a logo pipeline.
Markup edits make review loops practical, but unmarked pixels can move
Markup-guided editing is available on supported 2.5 and 2.6 models through the Image API. You send an annotation layer that marks the region to change, and you describe the change in the prompt. A reviewer circles a product detail, types the fix, and the model works on that region. That's a workflow people can run.
The catch is stated plainly: unmarked pixels aren't guaranteed to be preserved exactly. So review has two halves. Check the requested change, then compare everything outside the marked region with the original. Reject edits that drift past a tolerance you pick. It's dull work, which is what you want from a diff job. It fits well inside an n8n or Make flow where a person approves before anything publishes.
What we would do this week
Nothing here needs rushing. If you already generate images through the API, a small trial is cheap and tells you more than any launch post.
- Set
complexityexplicitly on every 2.6 call, starting at low for edits. - Run 20 of your real prompts through both Flare and Sunburst and compare cost per accepted asset.
- Sanitize every SVG before it reaches a page, and test GIF size and playback.
- Add a comparison of the area outside the marked region to any edit pipeline.
- Read OpenAI's pricing page before you commit to a budget.
Sources
- Changelog | OpenAI API · OpenAI
- AI Dev Weekly #33: Kimi K3.1 Pro Weights, Tencent Hy4 and GPT Image 2.6 · AI Made Tools
- OpenAI Ships Images 2.6, Raises Security Bug Bounties and Acquires Pierre for Codex · Scott Harvanek
Want this working in your stack, not just in a demo?
We are two senior engineers who build ai workflow automation with a fixed price per written scope.