Contra Labs and BenchFlow’s Frontier Design Benchmark evaluates AI on complex design work drawn from real client briefs in Contra’s marketplace. Professional designers compare model outputs without seeing their identities, while a separate critique rubric and catalog of common “AI slop” patterns provide more specific assessments than preference votes alone. The benchmark currently covers 33 briefs, 68 professional designers, 4,311 blind votes, and 35 rubric dimensions. Its organizers report that rubric scores correlate strongly with designers’ ratings (r = 0.94), and the gallery pairs client-approved “golden” designs with model-generated alternatives.
The leaderboard uses Elo ratings derived from head-to-head votes. Gestalt, labeled a stealth model, leads at 1,931 Elo with a 90% win rate; the human-created golden designs rank second at 1,739. Claude Fable 5.2 is third at 1,713, followed closely by Claude Opus 5.5, GPT-6.1 Sol, Claude Sonnet 5.5, and Kimi K3.1 Pro. The rankings therefore show Gestalt well ahead of Fable, while the supplied leaderboard does not identify a second stealth entry. Performance drops sharply below the top group: Qwen3.8 Max and several other models have win rates around 20-41%, and DeepSeek V4 Pro is last at 12%. The benchmark also offers its briefs, outputs, methodology, and rubric for inspection.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.