Gestalt AI中文

Frontier Design Benchmark methodology explained

Last updated:

Rather than one-line toy prompts, the Frontier Design Benchmark tests 33 commercial briefs—27 adapted from real Contra client commissions and 6 from startup partners—judged double-blind by 35 professional designers (median 8+ years of experience).

Real-world commercial briefs instead of toy prompts

Earlier UI benchmarks often relied on one-line prompts such as “build a coffee shop landing page.” Frontier Design Benchmark compiles 33 detailed briefs—27 adapted from real commercial commissions on Contra and 6 from startup partners.

Tasks span marketing websites, landing pages, web applications, and mobile interfaces, each carrying realistic brand assets, content constraints, and multi-section requirements.

Three evaluation pillars

To pair expert judgment with explainable diagnostics, Contra Labs and BenchFlow combine blind human preference voting with two structured scoring layers.

MechanismScale / metricWhat it measures
Blind Elo voting35 designers (8+ median yrs), 4,311 votesHead-to-head human preference without revealing model or human authorship
Design Critique Rubric35 analytical dimensions (r = 0.94)Structure, visual craft, content clarity, and usability (Gestalt 83, Claude Fable 5.2 84, Human 74)
Slop Detector0–100 AI-likelihood scoreVisual tells of generic AI output such as repetitive card grids and weak hierarchy

Reading leaderboard snapshots

Live Elo ratings update as additional blind votes are counted—for example, Gestalt moved from 1,928 Elo in the November 18, 2026 research report to 1,931 Elo in the November 21 snapshot.

Blind A/B evaluations focus on deliverable quality for a given brief rather than multi-week client iteration, accessibility testing, or design-system handoff.

FAQ

How many designers judged the Frontier Design Benchmark?

The published report records 4,311 blind head-to-head votes from 35 professional designers with a median of 8+ years of industry experience.

What is the Slop Detector?

It is a scoring engine built into the benchmark that estimates AI-likelihood on a 0–100 scale by flagging common AI design tells like generic card grids, predictable typography, and awkward spacing.

Why does the November 18 report show 1,928 Elo while the live table shows 1,931?

The research report reflects a fixed point-in-time snapshot from November 18, 2026, while the online leaderboard updates as additional blind votes are counted.

Gestalt AI

Model overview