Home / Blog / Two stealth models just appeared on the leaderboard
Tech News

Two stealth models just appeared on the leaderboard

Contra Labs and BenchFlow launched the FrontierDesign Benchmark, where stealth model Gestalt took first place with a 1931 Elo rating across 4,311 blind votes.

By Dillip Chowdary • Nov 19, 2026 • Source: contralabs.com

Two stealth models just appeared on the leaderboard
Image: contralabs.com

At a glance

Benchmark
FrontierDesign Benchmark
Collaborators
Contra Labs and BenchFlow
Dataset scale
33 client briefs and 4,311 blind votes
Evaluation panel
68 professional network designers
Top model
Gestalt stealth model with 1931 Elo
Rubric metrics
35 evaluation dimensions with r = 0.94

Contra Labs and BenchFlow launched the FrontierDesign Benchmark, an evaluation framework that measures artificial intelligence design capabilities using real-world client briefs from Contra's commercial marketplace. According to contralabs.com's report, the testing framework pits leading frontier systems against human designers in blind head-to-head comparisons scored by practicing design professionals. An unannounced stealth model named Gestalt secured the top spot on the leaderboard with an Elo rating of 1931, outranking human benchmark baselines and commercial frontier models.

This report details the complete benchmark standings, evaluation procedures, golden design gallery, and structured critique rubrics used to analyze model performance. The release provides creative directors, machine learning researchers, product designers, and engineering leads with quantitative data on how frontier generative models execute complex production briefs. The findings reveal noticeable performance tiers between frontier systems, highlighting where automated layout, composition, and visual polish meet professional standards.

01Blind head-to-head votes and Elo ratings

The benchmark rankings stem from 4,311 blind head-to-head comparisons evaluated by professional designers. Gestalt captured first place with an Elo rating of 1931 and a 90 percent win rate, maintaining a 95 percent confidence interval between 86 and 92 percent. Human golden designs finished second with an Elo of 1739 and a 74 percent win rate, edging out Anthropic's Claude Fable 5.2 at 1713 Elo and Claude Opus 5.5 at 1691 Elo.

ModelRankElo RatingWin Rate95% CI
Gestalt (stealth)#1193190%86–92%
Human golden#2173974%71–78%
Claude Fable 5.2#3171373%69–76%
Claude Opus 5.5#4169173%69–76%
GPT-6.1 Sol#5166870%66–74%
Claude Sonnet 5.5#6166769%65–73%
Kimi K3.1 Pro#7166371%66–74%
GPT-6 Astra#8163866%61–70%
Qwen3.8 Max#9143439%35–44%
Kimi K3#10141540%35–44%
GLM-5.3 Prime#11140341%37–45%
Gemini 4 Carbon#12136636%32–40%
Muse Spark 1.3#13133531%27–35%
Muse Wonder#14124922%18–25%
Grok 4.7#15124723%19–26%
Gemini 4 Argon#16122120%17–24%
DeepSeek V4 Pro#17112112%9–15%

OpenAI placed GPT-6.1 Sol in fifth place at 1668 Elo, closely followed by Claude Sonnet 5.5 at 1667 Elo and Moonshot AI's Kimi K3.1 Pro at 1663 Elo. A substantial separation appears below GPT-6 Astra at 1638 Elo, with open and mid-tier systems dropping below 1450 Elo. DeepSeek V4 Pro anchored the bottom of the board at 1121 Elo with a 12 percent win rate.

02Four evaluation stages across real marketplace briefs

The benchmark methodology establishes an empirical pipeline designed around production deliverables rather than synthetic prompts. Sixty-eight professional designers from the Contra Labs network conducted the reviews, assessing model submissions against commercially commissioned briefs. The evaluation framework organizes work across four core structural phases to maintain consistent grading across diverse creative requests.

  1. Brief selection Contra Labs extracts thirty-three complex client briefs directly from its active marketplace network.
  2. Designer assignment Professional network evaluators inspect competing design outputs against client specifications and constraints.
  3. Blind voting Evaluators cast head-to-head preference decisions between side-by-side candidate outputs without model identities.
  4. Rubric evaluation Reviewers grade submissions across thirty-five distinct analytical dimensions alongside an automated slop detector.

By relying on paid marketplace briefs, the benchmark tests how models interpret complex brand guidelines, information hierarchies, and responsive layouts. The sixty-eight network designers evaluated real deliverables rather than abstract concepts, grounding the results in commercial utility. This structure prevents models from exploiting narrow prompt patterns that fail in client-facing design workflows.

03Client briefs and the golden gallery

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

Contra Labs compiled a reference collection named the Golden Gallery, matching thirty-three client briefs against curated golden standard designs. Featured projects include website briefs for Bad Move, Bandstand, Basin, Broomwagon, Carafe, Compadre, Good Creature, NovaraAI, Off You Go, Ruckus, Threatlane, and Vellum. Each brief displays fifteen candidate model outputs, with Basin evaluating fourteen versions.

Commercial frontier models generated varying numbers of top-tier candidates across the catalog of briefs. Claude Fable 5.2 delivered primary outputs for Bad Move, Ruckus, and the NovaraAI marketing site. Claude Sonnet 5.5 produced outputs for Broomwagon, Off You Go, and Vellum, while Claude Opus 5.5 powered the Threatlane website delivery.

Alternative frontier architectures also earned positions across the golden catalog briefs. GPT-6.1 Sol generated submissions for Basin, Compadre, and Good Creature. Kimi K3.1 Pro provided outputs for Bandstand and Carafe, demonstrating competitive visual structuring against Anthropic and OpenAI entries across digital product categories.

04Thirty-five rubric dimensions and slop patterns

To supplement pairwise preference votes, the researchers established a design critique rubric covering thirty-five distinct dimensions. These rubric scores achieve a Pearson correlation coefficient of r = 0.94 with the designers' overall model ratings. This strong statistical alignment allows the benchmark to provide explainable qualitative assessments on top of standard Elo rankings.

The rubric pairs granular design criteria with a specialized detector targeting common AI visual flaws and slop patterns. Evaluators inspect typography choices, structural alignment, asset repetition, and element balance across individual components. Component tags in the dataset catalog specific layout fragments including frames, blooms, roses, sleeves, hands, sunflowers, buds, peaches, cherries, peonies, poppies, and marigolds.

The dual scoring system ensures that high Elo scores reflect genuine design craftsmanship rather than superficial visual appeal. By cross-referencing subjective pairwise choices against the thirty-five rubric criteria, evaluators separate coherent UI architecture from unfunctional aesthetic decorations. The framework isolates common generative artifacts that degrade responsive websites and client deliverables.

05Blind pairwise votes and side-by-side outputs

The voting environment presents two competing model outputs for the same client brief simultaneously without exposing model identities. Human evaluators interact with the live designs, inspect structural details, and either select the superior implementation or record a tie. The testing interface incorporates responsive controls that allow reviewers to drag elements and inspect fine typography.

The blind testing methodology eliminates brand recognition bias, forcing models to compete strictly on output quality. Evaluators judge how well each model resolves complex typographic hierarchies, navigation menus, and content grids under identical prompt conditions. The aggregated pairwise choices feed directly into the Elo calculation system to produce the overall rankings.

Note: Gestalt remains an unannounced stealth model on the public leaderboard, with no commercial release date, pricing tier, or developer identity disclosed by Contra Labs or BenchFlow. The benchmark organizers have not confirmed when the weights or API will become publicly accessible.

Frequently asked questions

What is the FrontierDesign Benchmark?

It is an evaluation benchmark developed by Contra Labs and BenchFlow to test AI model design capabilities against real marketplace briefs. The system uses blind head-to-head votes from professional designers and thirty-five rubric dimensions.

Which model ranks highest on the FrontierDesign leaderboard?

Gestalt, an unannounced stealth model, ranks first overall with an Elo rating of 1931 and a 90 percent win rate. It ranks ahead of human golden designs at 1739 Elo and Claude Fable 5.2 at 1713 Elo.

How closely do design rubric scores track human designer preferences?

Rubric scores correlate with professional designers' ratings at a coefficient of r = 0.94. This correlation provides structured, explainable design quality measurements alongside pairwise blind preference votes.

How many design projects and votes are included in the dataset?

The benchmark incorporates 33 real client briefs, evaluations from 68 professional designers, 4,311 blind head-to-head votes, and 35 rubric dimensions.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →