We Benchmarked GPT-Image-2.6 Native <svg> Output: Path Counts, Cost & Real Edge Cases

A developer benchmarked gpt-image-2.6's native output_format: "svg" support while building GPTSVG Studio, finding that native LLM-authored SVG geometry produces compact files (0.9 KB to 9.6 KB, 5 to 28 paths) with grouped layers and no embedded raster bitmaps, unlike auto-traced raster output. The benchmark also found native SVG output is restricted to text prompts, with image-edit endpoints rejecting the format, and the developer open-sourced the sample outputs and a zero-dependency TypeScript XML AST sanitizer under MIT, warning that LLM-generated SVG should never be injected via innerHTML.

Most "AI vector generators" over the past two years did the exact same thing under the hood: generate a raster PNG, then run an aggressive auto-tracer over it—leaving you with 1,800 jagged <path nodes and muddy gradients that bog down Figma. With gpt-image-2.6 adding native output format: "svg" support, we ran a structured benchmark while building GPTSVG Studio https://gptsvg.com to see whether native LLM-authored <svg geometry is actually clean enough for production frontend and brand work. All raw .svg sample outputs, rendered previews, and our standalone zero-dependency TypeScript XML AST sanitizer are open-sourced under MIT here: 👉 GitHub Benchmark & Sanitizer Repo : https://github.com/keokeo/gpt-image-2-6-svg-benchmark https://github.com/keokeo/gpt-image-2-6-svg-benchmark <g id="..." layers—not traced soup 0.9 KB – 9.6 KB , zero embedded data:image/png bitmaps, and readable group IDs: | Preset | Quality | Raw Size | Elements / Paths | Embedded Raster | | :--- | :--- | :--- | :--- | :--- | | Minimal Geometric Logo | medium | 1.4 KB | 6 elements | 0 | | Wordmark + Mark | medium | 2.1 KB | 9 elements | 0 | | 24×24 Monoline UI Icon | low | 0.9 KB | 5 strokes | 0 | | Die-Cut Sticker | medium | 9.6 KB | 28 paths | 0 | | Shield / Crest Badge | low | 1.9 KB | 8 paths | 0 | Native output format: "svg" is strictly Text-to-SVG Upstream providers currently restrict native SVG output to text prompts; image-edit endpoints reject output format: "svg" . For Image to SVG conversion https://gptsvg.com/image-to-svg , we paired a zero-upload client-side Web Worker k-means vectorizer free, runs 100% locally in your browser with an optional flat-palette AI redraw pass. Never inject LLM-generated SVG with innerHTML Inline <svg is executable XML. Treat model-generated SVG as untrusted input: parse it into an XML AST, rebuild only allow-listed geometry tags/attributes stripping <script , <foreignObject , on handlers, and external href references , serve stored files with Content-Security-Policy: default-src 'none'; sandbox , and render previews via <img src="blob:..." . Happy to share raw SVG snippets or sanitizer edge cases if anyone else is building vector tooling