Newshot

Anonymous AI Model Gestalt Tops Contra's Frontier Design Benchmark: Far Outperforms Claude Fable 5.2 and GPT-6.1 Sol, Elo Score Beats Human Reference Works

AI summary

Contra Labs released its Frontier Design Benchmark for AI design capability on November 18 local time. An anonymous AI model codenamed Gestalt, currently in stealth development, topped the blind test by professional designers by a clear margin, surpassing mainstream models such as Anthropic's Claude Fable 5.2 and OpenAI's GPT-6.1 Sol, and even outranked human designers' reference works in designer preference scores. According to the official results, Gestalt ranked first with an Elo score of 1931 and a 90% win rate.

Translation(model translation, not the source's own text)

Anonymous AI Model Gestalt Tops Contra's Frontier Design Benchmark: Performance Far Exceeds Claude Fable 5.2 and GPT-6.1 Sol, Elo Score Surpasses Human Reference Works

IT Home, November 21 news: Contra Labs released the Frontier Design Benchmark, an AI design capability evaluation leaderboard, on November 18 local time.

Among them, an anonymous AI model codenamed Gestalt, currently in stealth development, became the biggest dark horse, topping the blind test by professional designers with a clear margin, surpassing mainstream models such as Anthropic's Claude Fable 5.2 and OpenAI's GPT-6.1 Sol, and even surpassing the reference works of human designers in designer preference ratings.

According to the evaluation results officially published by Contra Labs, Gestalt achieved an Elo score of 1931 with a win rate of 90%, ranking first on the leaderboard.

Ranking second is the human designer reference work "Human golden," with an Elo score of 1739 and a win rate of 74%; Claude Fable 5.2 ranks third with 1713 points, while GPT-6.1 Sol ranks fifth with 1668 points.

This evaluation was launched by Contra Labs in collaboration with BenchFlow and focuses on real commercial design tasks rather than simple prompt-generation tests. There are 33 evaluation tasks in total, of which 27 are adapted from real design commissions on the Contra platform, and the other 6 come from collaborative projects with startups. The evaluation invited 35 professional designers to participate and collected a total of 4,311 votes through 5 rounds of blind testing.

The evaluation uses three metrics: double-blind preference voting by professional designers, a design quality evaluation system (Design Critique Rubric), and an AI trace detector (Slop Detector). Among them, the Elo score is calculated based on designers' pairwise comparisons of different works, covering color, typography, layout, usability, and adherence to task requirements.

In direct matchups between Gestalt and human designer reference works, designers chose Gestalt in 20 out of 28 comparisons. However, this does not mean that Gestalt comprehensively surpasses human works on all design quality metrics.

In the design quality evaluation system, Gestalt scored 83 points, Claude Fable 5.2 and Claude Opus 5.5 both reached 84 points, while the human designer reference works scored 74 points. Contra Labs noted that designers' aesthetic preferences are not entirely the same as specific execution quality; although the human reference works have some clearly identifiable execution issues, they still received relatively high preference votes in the blind test.

In terms of AI trace detection, Gestalt's works also performed prominently, with the AI trace score of its generated results approaching that of the human reference works. The detector uses a classification model to determine how many typical features of AI-generated works a design exhibits; a lower score means the work looks less like it was AI-generated.

Currently, Contra Labs has not disclosed the development team, model architecture, or specific identity of Gestalt, marking it only as Stealth. The actual origin of the model and the team behind it remain to be further confirmed.

  • "The World's Strongest AI Model: Claude Fable 5.2 Arrives, Outperforms GPT-6 Astra in Coding and Tops Multiple Benchmarks"

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗