In Introducing Images 2.6, OpenAI states that SVG export lets designs go straight into websites, apps, and print. The API Changelog (November 4, 2026) lists SVG generation as a formal capability of GPT‑Image‑2.6 Flare and Sunburst. The image generation guide specifies requesting vector output by setting output_format to "svg" on POST /v1/images/generations; uses are summarized as graphics that need editable paths, shapes, and colors—icons and illustrations, for example. This article only states documented boundaries—no invented print-resolution guarantees or cross-vendor compatibility scores.

👉 Try the GPT Image 2.6 API on PoYo: per-image pricing and online playground.

Request fields and hard limits

Per the image generation guide and image prompting guide, SVG requests must follow these constraints. First, n must be 1—no multi-vector batch return. Second, streaming, partial images, output_compression, and prompt_rewriter are unsupported. Third, the image edits endpoint does not support SVG output; SVG is also not an acceptable raster input to edits, so “emit SVG then edits paths” is not an official path. Fourth, the API generates a reference image first and derives SVG from it; the quality tier applies to that reference; response b64_json should be base64-decoded and saved as UTF-8 XML with a .svg extension. Fifth, the returned document is static vector content—no scripts, external resources, or embedded rasters. SVG also supports background="transparent" for placing marks on arbitrary backgrounds.

Both Flare and Sunburst can generate SVG; the prompting guide compares both tiers on the same “Field & Flour” bakery-logo prompt, showing that selection can still follow the usual speed-vs-precision logic rather than assuming one tier is “more vector.” Billing follows model-page image-token rules; complexity changes have no separate fee. Measure per-item cost on representative logo and icon prompts before launch.

Fit use cases: web, app, and print-bound

The product announcement positions SVG as export that “lets designs move into the next stage of a project.” Combining that with doc language, durable fit statements include: logos and icons that need path-level re-editing; flat illustrations that must scale losslessly on web and clients; brand marks that need a transparent background for color overlays; and pre-press vector drafts that should still allow color and stroke edits in design tools. On prompting, official examples stress clean outlines, strong silhouettes, balanced negative space, and few gradients so marks stay readable at small and large sizes—aligned with common print and app-icon acceptance bars, but not a claim of “tested on a specific press.”

For web and app pipelines, returned XML can go to frontend or design-system cleanup; docs do not claim the SVG always meets a specific optimizer or accessibility profile—integrators should run their own structure checks, path merges, and sanitization. For print-bound workflows, vectors avoid raster upscale jaggies, but color management, spot colors, bleed, and font embedding remain production-system duties outside Image API docs. If the final must be editable raster or motion, switch to PNG/WebP/GIF rather than forcing SVG as a universal container.

Boundaries not to overstate

First, SVG is a generations-side capability—do not expect edits or markup to emit vectors directly. Second, returns contain no scripts or external resources—do not describe them as “interactive SVG apps.” Third, docs explicitly disallow SVG as edit input, so brand iteration should happen in the raster stage with references, markup, or mask, then request SVG export separately after freeze. Fourth, general limits still apply: complex prompts can take up to about two minutes; text rendering and structured composition may still fall short. Fifth, content moderation and the moderation parameter apply to SVG requests as well.

For engineering acceptance, build a fixed prompt set: monochrome logos, multicolor icons, lockups needing transparent backgrounds, and marks with a little necessary text. After decode, check that the root element and paths are complete, that scripts or external links were not wrongly included, and that transparent regions are truly hollow. Run your own SVG optimize and sanitize flow on web/app; hand color profiles, bleed, and spot colors to prepress systems rather than expecting the Image API to do that work. The quality tier affects the reference raster and thus can indirectly affect how fragmented vector paths become—if paths are too fragmented or too coarse, raise or lower quality with other conditions frozen rather than changing prompt and tier together. Flare suits fast shape trials; Sunburst suits pre-final assessments that are stricter on outline and detail; both can emit SVG.

Bottom line: Images 2.6 puts static SVG on the Image API’s formal output list for vector reuse before web, app, and print. Integrate with output_format: "svg", n=1, and the hard rule “generations only, no edits-side output,” and complete path acceptance and prepress in your own toolchain without overstating capability.