Design Arena
Crowdsourced human-preference benchmark for design and code-generation outputs. Use it as directional taste/eval evidence, not as a single source of truth for model routing.
Design Arena presents blind, pairwise model outputs to users and ranks models with Elo ratings from votes. Its own site describes the platform as a global crowdsourced benchmark for design with head-to-head model comparison, 5M+ votes, and 30+ categories spanning websites, UI components, games, 3D design, data visualization, SVG, image generation, video, audio, world models, and agentic web/app development. Source: Design Arena site, 2026-06-30
Why It Matters
Design work is hard to evaluate with static unit tests. Human-preference arenas give a noisy but useful proxy for taste, instruction-following, aesthetics, and final artifact appeal. The strongest use is relative evidence: "model A is currently winning this category under this judge population," not "model A is objectively best for our product."
For Kevin's stack, Design Arena belongs in the same evaluation family as Agent Self-Improvement Eval Library and Red Queen Gödel Machine: it reminds agents to make evaluators harder and more representative, especially for design and agentic web-dev work where screenshots and live artifacts matter.
GLM-5.2 Snapshot
The reviewed bookmark's chart shows GLM-5.2 at 1360 Elo and #1 in a Code Categories Arena snapshot, ahead of Claude Fable 5 (1350) and Claude Opus 4.6 (1342). The tweet says GLM-5.2 improved four positions and 27 Elo points, and notes that "agentic evaluations" were coming soon. Preserve this as a dated benchmark snapshot, because live leaderboards move. Source: X/@Designarena, 2026-06-16
Routing
- Check Design Arena when selecting models for visual UI generation, agentic web development, design-to-code, and code-category artifact quality.
- Pair arena rankings with project-specific evals. A public Elo ranking cannot replace a repo-local task fixture, screenshot diff, accessibility check, or product taste review.
- Record date, category, Elo, and source image when using a leaderboard claim. Benchmark claims rot quickly.
Timeline
- 2026-06-30 | Created from deep review of Kevin's Design Arena bookmark, local chart artifact, and live Design Arena site. Source: X/@Designarena, 2026-06-16; Source: Design Arena, 2026-06-30