SVG-Bench

SVG-Bench is Datapoint AI’s open human-preference dataset and public leaderboard for text-to-SVG generation, announced on October 22, 2026. Datapoint calls the board Datapoint SVG Bench: a public evaluation of language models that write SVG from text prompts, ranked from head-to-head human votes on the rendered images. The dataset holds 352,270 validated pairwise votes, exactly 10 for each of 35,227 SVG pairs.

Fourteen models wrote SVG for 400 original prompts, 50 in each of eight categories: icons, logos, illustrations, scenes, diagrams, typography, UI, and precision. 20,323 annotators in 183 countries saw two Chromium renders at 1024×1024 pixels on white, plus the prompt, and answered “Which image do you prefer?” Model names were hidden. Every output is included as SVG source and as the PNG annotators saw, with token use, list-price cost, and latency. The overall Elo board gives each category an equal share and matches the live leaderboard.

GPT-6 Astra (OpenAI) ranks first at 1065.6 Elo with a 59.8% win rate, Gemini 3.8 Flash (Google) second at 1059.2, and Muse Wonder (Meta) third at 1047.3. Claude Fable 5.1 (Anthropic) is fourth at 1028.5. Arrow 2 Telos (QuiverAI) is eighth overall at 1011.6 and leads Illustrations and Scenes.

Votes, prompts, and metadata are CC-BY-4.0 with attribution to Datapoint AI. Annotators must not be re-identified. The SVGs and renders are third-party model outputs for research and evaluation, under each provider’s terms, and are not a curated training corpus for SVG generation.

Load it in Python with datasets.load_dataset("datapointai/text-to-svg-human-preferences-352k"). The live board is on trydatapoint.com and the dataset card is on Hugging Face.