Mystery AI Model Gestalt Tops Contra Labs Design Benchmark Above All Named Rivals
The stealth model reached 1,928 Elo — 214 points above Claude Fable 5.2 — and no lab has claimed credit

An artificial intelligence model that nobody has publicly claimed built has topped the first Contra Labs Frontier Design Benchmark, outscoring human professional designers, all named frontier models, and every other entrant in a blind pairwise evaluation conducted by 35 professional designers across 33 real commercial briefs.
The model, identified only as Gestalt, reached 1,928 Elo on the leaderboard — 214 points above Claude Fable 5.2, the next-highest named AI model, and 189 points above the human golden reference tier at 1,739. Human professionals, who set a high bar, scored distinctly above all public frontier models except Gestalt. Gestalt scored above them too, winning 20 of 28 head-to-head matchups against human reference design work.
Contra Labs, the online freelance design marketplace that organized the evaluation with AI benchmarking platform BenchFlow, describes Gestalt as "a high-quality design model currently in stealth" and says it is "working with the team to continuously evaluate its quality and provide feedback." No AI laboratory, research group, or private developer has identified itself as the creator.
A Benchmark Built on Real Work — Not Test Prompts
The methodology separates the Frontier Design Benchmark from most AI evaluations, which typically rely on standardized test sets or expert-panel scoring. Contra Labs used 33 actual client briefs drawn from its marketplace — 27 sourced from real Contra projects and 6 from startup engagements. Each AI model generated design outputs from the same briefs. The 35 professional designers then voted in blind pairwise comparisons without knowing which model produced which output.
Across the full evaluation, 4,311 pairwise votes were cast over five rounds. The Bradley-Terry pairwise model converted those votes into Elo ratings. To measure how reliable those ratings were, Contra computed a split-half Spearman reliability coefficient across 200 random splits of the voting data, yielding a coefficient of 0.95 — a figure indicating high internal consistency. The evaluation also incorporated a 35-dimension design rubric covering quality dimensions independent of pairwise preference, and found a correlation of r = 0.94 between rubric scores and vote outcomes.
How the Leaderboard Looks
The tier structure separating Gestalt from the rest of the field is wide. The leading cluster of named frontier models — Claude Fable 5.2 at 1,714 Elo, Claude Opus 5.5 at 1,691, GPT-6.1 Sol at 1,668, Claude Sonnet 5.5 at 1,667, and Kimi K3.1 Pro at 1,663 — spans only 51 Elo points. Gestalt sits 214 points above that group's top performer.
The second tier falls 204 Elo points below the first. DeepSeek V4 Pro ranked last in the evaluation with a score of 1,122, a result Contra attributed primarily to agent execution failures during task completion rather than raw design quality.
Human golden reference designs, compiled from the professional designers' own portfolio work, scored 1,739 Elo — well above the named frontier cluster, but 189 points below Gestalt.
The Slop Detector Finding
One of the benchmark's additional measurement tools is a binary classifier called the Slop Detector, trained to distinguish AI-generated design pages from human-made ones on a scale from 0 to 100, where lower scores indicate output more consistent with human production.
Gestalt scored 1.2 on the Slop Detector — almost identical to human professional designs, which scored 0.9. That result is a notable finding in its own right: by this classifier's measure, Gestalt's output is nearly indistinguishable from human design work. By comparison, Claude Fable 5.2 scored 31 on the same classifier, the GPT-family models scored in the 41–53 range, Gemini models scored 54–57, and Kimi K3.1 Pro scored 66.
Contra Labs reported the rubric score for human golden designs as 73.6 — comparatively low for designs that nonetheless ranked near the top of the preference leaderboard. The pattern suggests that professional designers, voting blind, found work compelling even when it deviated from formal rubric criteria, while Gestalt combined high preference performance with near-human classifier scores.
The SSI Rumor Was a Self-Admitted Fabrication
Shortly after the benchmark results circulated, speculation connecting Gestalt to Safe Superintelligence Inc. — Ilya Sutskever's AI safety laboratory — spread on social media. The claim was quickly identified as a self-admitted fabrication with no evidentiary basis. SSI has released no public AI model as of November 2026, and no sourced reporting has connected the laboratory to Gestalt.
Contra Labs has not publicly named or hinted at the developer's identity beyond confirming that it is in an active working relationship with the team.
What the Gap Signals
The Frontier Design Benchmark's key editorial observation is not simply that Gestalt won. It is that the gap between Gestalt and the established frontier is unusually wide by the standards of competitive AI evaluations, where leading models on most benchmarks are separated by small margins.
A 214-point Elo lead over Claude Fable 5.2 — itself 25 points ahead of Claude Opus 5.5, which is 23 points ahead of GPT-6.1 Sol — suggests Gestalt is not operating at the margin of the frontier but substantially above it, at least for the visual design tasks this benchmark measures.
Whether the same holds across a broader range of AI design tasks, or whether the benchmark's methodology introduces biases that favor certain kinds of AI output, remains open. For now, the benchmark is public, the methodology is documented, and the model at the top of it has no name attached to it.



