跳到正文
Hacker News· cheptsov·· 17 小时前AI 评分40

Contra Labs 设计基准榜出现两个匿名模型,排名均超过 Fable 5.2

Two stealth models just appeared on the leaderboard, both ahead of Fable 5.2

AI 导读

Contra Labs 的 Frontier Design Benchmark 排行榜上出现两个未具名模型,Elo 排名均高于 Claude Fable 5.2。该基准基于 33 个真实客户设计 brief、68 位专业设计师的 4,311 次盲投票,并配有 35 个维度的设计评审 rubric,rubric 评分与设计师评分的相关系数达 r = 0.94。

正文 · 原文

Home

Deconstructing Real World Creativity.

Exploring the frontier of AI design capabilities through real client briefs from Contra’s marketplace, golden designs and model comparisons.

© 2026 Contra Labs × BenchFlow

Leaderboard

How frontier models rank in blind head-to-head votes on real client briefs from Contra’s marketplace.

Contra Labs

Frontier
Design Benchmark

Elo rating from blind head-to-head votes

  1. Gestalt
  2. Human golden
  3. Claude Fable 5.2
  4. Claude Opus 5.5
  5. GPT-6.1 Sol
  6. Claude Sonnet 5.5
  7. Kimi K3.1 Pro
  8. GPT-6 Astra
  9. Qwen3.8 Max
  10. Kimi K3
  11. GLM-5.3 Prime
  12. Gemini 4 Carbon
  13. Muse Spark 1.3
  14. Muse Wonder
  15. Grok 4.7
  16. Gemini 4 Argon
  17. DeepSeek V4 Pro

contralabs.com/frontierdesign

Real design projects, judged by designers and scored by rubrics

Frontier Design Benchmark is the first benchmark built on real world, high value, and complex design projects. It is judged by professional designers in the Contra Labs network, ensuring the quality of evaluation.

In the benchmark, we also built a design critique rubric and a list of AI slop patterns. The rubric scores are highly correlated with the designers’ ratings of the models (r = 0.94). These additional metrics provide explainable design quality judgement on top of pairwise preferences.

33
design briefs
68
professional designers
4,311
blind votes
35
rubric dimensions

Golden Gallery

Real client briefs, each with its golden: the design set beside the models’ outputs as the bar.

View all

Sign inGet in touch

来源:Hacker News · contralabs.com