/
Leaderboard
Benchmarks

Leaderboard

Three benchmarks on real UI work: copying a component, building inside a product's design system, and getting picked by developers next to real designers.

1,400 copies280 design system runs~17,000 developers

Copy Bench

0-100

Models rebuild a web component from a few seconds of screen recording. We score how close each copy gets.

#Model · per copyScore
  1. 1GPT-6 Astra$8.70 · 22 min86.7
  2. 2GPT-6.1 Sol$1.41 · 20 min82.2
  3. 3Claude Fable 5.2$0.90 · 2.6 min69.0
  4. 4Gemini 4 Carbon~$1.84 · 43 min67.2
  5. 5GPT-6 Sol$1.27 · 6.9 min64.5
  6. 6Claude Opus 5.5$0.44 · 2.3 min59.1
  7. 7Qwen 4 Max~$0.90 · 24 min54.4
  8. 8Muse Wonder 1.0~$1.74 · 8.1 min51.9
  9. 9Qwen 3.8 Max Prime~$3.62 · 21 min50.1
  10. 10Kimi K3.1 ProOpen~$0.56 · 8.7 min46.6
Updated Nov 16All 14 models

Design System Bench

0-100 ± 95%

Coding agents build UI from a short ticket in a real product's repo. We score how often they use its own design-system components.

6 products20 tasks
#Model · per taskScore
  1. 1Claude Fable 5.2$0.99 · 1.4 min85.2±7.2
  2. 2Gemini 4 Carbon~$1.31 · 32 min82.9±10.0
  3. 3GPT-6 Astra$1.21 · 1.6 min79.9±9.2
  4. 4Qwen 3.8 Max Prime~$2.34 · 10 min78.8±8.5
  5. 5GPT-6.1 Sol$0.20 · 1.3 min78.5±9.1
  6. 6MiniMax M3Open~$0.68 · 12 min77.3±9.3
  7. 7GPT-6 Sol$0.31 · 1.2 min75.2±8.4
  8. 8Claude Opus 5.5$0.64 · 1.7 min73.7±7.7
  9. 9Grok 4.7$0.65 · 1.9 min72.7±7.6
  10. 10GLM 5.3~$0.89 · 6.4 min71.5±9.6
Updated Nov 12All 14 models

Turing UI Bench

pick rate

AI-made hero sections shown unlabeled in the 21st feed next to work by real designers. We count which ones developers open.

12 models12 designers
#Model · for 4 heroesPicked
  1. ·Top designer4.7%
  2. 1GPT-6 Astra$8.09 · 17 min3.4%2.8-4.2
  3. 2GPT-6.1 Sol$1.50 · 14 min3.3%2.7-4.1
  4. 3Grok 4.7$2.41 · 22 min2.8%2.2-3.5
  5. 4Claude Sonnet 5.5$1.71 · 6.0 min2.7%2.1-3.3
  6. ·Median designer2.7%
  7. 5Claude Fable 5.1$15.96 · 17 min2.6%2.1-3.3
  8. 6Muse Spark 1.3$0.41 · 2.0 min2.6%2.1-3.2
  9. 7Gemini 3.8 Flash$0.22 · 1.5 min2.5%2.0-3.1
  10. ·Lowest designer1.1%
Oct 13 to Nov 2All 12 models

How each benchmark works, the research behind it and lab partnerships: All benchmarks