Benchmarks
357 sources · 1,263 models × 3,079 benchmarks · 162,127 scoresContra Labs' Frontier Design Benchmark (launched November 18, 2026, with BenchFlow) gives AI agents 33 website-design briefs, 27 rewritten from real design jobs posted on Contra for fictional clients and 6 from startups, each with the client's brand files. BenchFlow ran each model as an agent in a sandbox through OpenRouter, in its own lab's agent harness wherever it could. Professional designers on Contra vote blind on pairs of outputs for the same brief, with names hidden and sides shuffled, and the votes are fit with a Bradley-Terry model reported on the Elo scale; a model is ranked after 10 counted votes. The pool also rates human designers' sites, built from mood boards with any tools: at launch they placed second at 1739, above every named model. Every vote refits the pool, so this is a diagnostic.
No signal or difficulty estimate. Held out of the fit by design.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Fable 5.2 | not reported | 1713 | |
| 2 | Claude Opus 5.5 | not reported | 1691 | |
| 3 | GPT-6.1 Sol | not reported | 1668 | |
| 4 | Claude Sonnet 5.5 | not reported | 1667 | |
| 5 | Kimi K3.1 Pro | not reported | 1663 | |
| 6 | GPT-6 Astra | not reported | 1638 | |
| 7 | Kimi K3 | not reported | 1415 | |
| 8 | GLM-5.3 | not reported · GLM-5.3-Prime inference-speed tier through OpenRouter | 1403 | |
| 9 | Gemini 4 Carbon | not reported | 1366 | |
| 10 | Muse Spark 1.3 | not reported | 1335 | |
| 11 | Muse Wonder | not reported | 1249 | |
| 12 | Grok 4.7 | not reported | 1247 | |
| 13 | Gemini 4 Argon | not reported | 1221 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).