Brightline
Archivo · 6px corners
Which AI designs on brand?
Every model gets the same brand file and the same screens to build. A real browser renders each page, and designlang measures how close it stayed to the brand and how generic it looks. Measured, not judged.
The table
Run on 4 November 2026: 5 brands × 4 screens per model. On brand is higher-is-better; generic look is lower-is-better.
| # | Model | On brand | Generic look | Failed runs |
|---|---|---|---|---|
| 1 | Gemini 4 Argon | 94 | 4 | 2 of 20 |
| 2 | Claude Sonnet 5 Anthropic | 93 | 1 | 0 of 20 |
| 3 | Kimi K3.1 Pro Moonshot AI | 92 | 2 | 1 of 20 |
| 4 | GPT-6.1 Sol OpenAI | 91 | 3 | 0 of 20 |
| — | GLM-5.3 Zhipu AI | 95 | 6 | 16 of 20 Not ranked: most runs returned nothing |
A failed run (no reply, something other than one complete HTML page, or a page that would not render) counts against its model and is left out of the averages. Runs the provider dropped were tried once more. A model is ranked only when at least three in four of its runs completed. Models were called through Atlas Cloud; the run cost $14.25.
The brands
Five fictional brands, each a DESIGN.md with colours, fonts, corners and a spacing grid. None of them is a real company.
Ember
Bricolage Grotesque · Source Serif 4 · 0px corners
Fernwood
Fraunces · Work Sans · 4/10px corners
Halcyon
DM Serif Display · Nunito Sans · 12/24px corners
Koto
Space Grotesk · IBM Plex Sans · 0/2px corners
Screens: Landing page hero, Pricing page, Dashboard, Settings.
Run it yourself
The briefs, prompts, runner and scorer are open source. Bring your own key and a spending cap.