Sulat.com / Writing on AIFriday, November 20, 2026
Artificial Intelligence/19 November 2026

Contra Labs’ Frontier Design Benchmark puts stealth model Gestalt ahead of human designers

Contra Labs' new Frontier Design Benchmark has designers vote on AI-built sites from real client briefs. A stealth model, Gestalt, leads at 1,928 Elo and beat human golden designs in 20 of 28 matchups.

Glossy 3D Contra Labs bookmark badge floating above a podium of wireframes and colour swatches

At a glance

  • Contra Labs' Frontier Design Benchmark has professional designers vote blind on AI-built sites from 33 design briefs.
  • Stealth model Gestalt leads at 1,928 Elo and was picked over a human golden design in 20 of 28 matchups.
  • A 204-point gap separates the top group of models from the rest.
  • The design rubric ranks human golden designs sixth, but designers chose them in 74% of votes.

Contra Labs has launched the Frontier Design Benchmark, a leaderboard where professional designers vote on websites and apps built by AI models. A model called Gestalt, which is still in stealth, sits first at 1,928 Elo. Contra Labs says it’s the first model rated above human golden designs. Ben, who builds Contra and Contra Labs, announced the result on X, calling Gestalt a model that’s beating everything, including the human golden designs. Contra Labs hasn’t named the lab behind Gestalt.

How the benchmark works

The benchmark has 33 design tasks. Twenty-seven come from real jobs posted on Contra, rewritten for fictional clients, and six come from projects that startup partners hired a Contra designer to do. Contra Labs looked at 4.6 thousand website jobs from the past year and found the median brief runs 138 words on a US$1,050 budget, with 75% stacking two or more constraints such as branding assets or business goals. The tasks range from a festival ticketing site to a museum audio tour app.

BenchFlow ran an AI agent on each task in a sandbox and recorded the output, cost and full trajectory. Models went through OpenRouter, and each ran in its own lab’s agent harness where that was possible. For the human side, Contra designers built golden designs for each task, starting with a moodboard of three directions and then building the site with tools of their choice.

Four stat cards: 33 briefs, 4,311 votes, 35 designers, Gestalt preferred over golden designs 20 of 28 times
Designers chose Gestalt over a human golden design in 20 of 28 head-to-head comparisons. Source: Contra Labs.

Gestalt leads by a wide margin

Designers cast 4,311 blind votes over five rounds, and Contra Labs turned them into Elo scores with a Bradley-Terry model. Gestalt wins 89% of its votes (378 of 423). The best model Contra Labs actually ran is Claude Fable 5.2 at 1,714, which puts Gestalt 214 points ahead with no overlap in their 95% intervals. When designers saw a Gestalt site next to a golden design, they picked Gestalt 20 times out of 28.

The rest of the field falls into two groups. Six models sit between 1,638 and 1,714 Elo: three Claude models, Kimi K3.1 Pro and two others, each winning 66% to 73% of its votes. The other nine range from 1,122 to 1,434 and win 12% to 41%. That 204-point gap between the groups is more than twice the spread inside the top one.

Dot chart: Gestalt 1,928 Elo, top group near 1,700, lower group near 1,400, a 204-point gap between groups
Gestalt leads at 1,928 Elo, and a 204-point gap splits the other models into two groups. Source: Contra Labs.

Rubric scores and designer taste disagree

Contra Labs also grades every output against a design rubric of common mistakes, such as too many fonts or uneven spacing. Across 15 models the rubric score tracks Elo closely, with a correlation of 0.94. The human golden designs are where the two part ways. The rubric puts them sixth of 16 with 73.6 points, behind five models, because it docks execution slips like a button label too dark to read or a wordmark cut off by the section above. Designers saw those same slips and still chose the golden designs in 74% of votes (425 of 574).

Two cards: rubric ranks golden designs 6th of 16, designers chose them in 74% of votes
The rubric ranks human golden designs sixth, yet designers picked them in 74% of their votes. Source: Contra Labs.

What counts as AI-looking

The third measure is a Slop Detector, a classifier trained on human-made and AI-made pages that reads from 0 to 100. Human golden designs read 0.9. The models run from 1.2 to 66, and the three Claude models read 31 to 38. Two entrants read 1.2 for opposite reasons. Many of DeepSeek V4 Pro’s pages are unfinished, so they don’t look like typical AI pages, while Gestalt’s pages are finished and designers still picked them in 89% of votes. Contra Labs says Gestalt’s output is almost on par with human golden designs on this measure.

The detector comes with a list of tells that designers flag. A small spaced-out capital label above a heading shows up on at least half the sections of 93% of AI pages in its training set, against 11% of human pages. Numbered sections such as 01, 02, 03 appear on 17% of AI pages and on no human page. An italic word in the headline turns up on 60% of Kimi K3.1 Pro’s pages. Claude models put a horizontal arrow on links and buttons, on 88% of Fable 5.2’s pages, against 49% for other labs.

What comes next

Contra Labs says the 33 tasks are an initial set and that it will add more real briefs, models and design artifacts. It also says it’s working with the Gestalt team to keep evaluating the model. The leaderboard, gallery and a public vote page are on the Frontier Design site, and Contra Labs ran the whole benchmark with BenchFlow.

Hero illustration features the Contra Labs logo, which remains the trademark of its owner.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.