A model named Gestalt is leading a new AI design leaderboard. But its maker has not been named, the model is not publicly available, and the result comes from one benchmark run by the company that introduced it.

That does not make the result useless. It does mean we should separate what the test measured from what it cannot tell us.

What is Gestalt?

Gestalt is the name Contra Labs gives to a design model it describes as being in stealth. Contra has not publicly identified the team behind it or released technical details such as its architecture, training data, or model size. The model is not available for readers to try, based on the public pages reviewed for this article.[1][2]

So the headline is not that a new public model has arrived. It is that an unnamed system scored unusually well in a design evaluation hosted by Contra Labs and BenchFlow.

What the benchmark tested

The Frontier Design Benchmark asks models to build websites and apps from design briefs. The first task set contains 33 briefs. Contra says 27 were adapted from real jobs posted on its marketplace, with client details rewritten, while six came from startup partners.[1]

Professional designers compared two outputs for the same brief without being told which came from a model or a person. Contra used those blind choices to calculate Elo ratings, a way to rank competitors from their results against one another. The live leaderboard reports 4,311 votes across the benchmark.[2]

The evaluation also includes a design rubric and a tool Contra calls its Slop Detector. The rubric checks design details such as spacing and type hierarchy. Contra reports that rubric scores correlate with designer rankings at r = 0.94. That is a strong relationship in this dataset, but it does not mean the rubric captures every part of good design.[1]

What Gestalt's lead means

In its November 18 launch report, Contra put Gestalt at 1,928 Elo, with 378 wins from 423 votes. It said designers chose Gestalt over a human reference design in 20 of 28 comparisons.[1]

The live leaderboard later showed Gestalt at 1,931 Elo, with a 90% win rate across 420 votes. Human reference designs ranked second at 1,739, followed by Claude Fable 5.2 at 1,713.[2] These are different snapshots, so the figures should not be mixed as if they came from the same moment.

The results support a limited conclusion: in this benchmark, designers preferred Gestalt's submitted pages over the alternatives they saw. The direct comparison with human work is worth attention, but 28 votes is a small slice of the total. It cannot establish that Gestalt is better than human designers across the wider range of design work.

Contra also says Gestalt's outputs looked less AI-generated to its Slop Detector, close to its human reference designs.[1] That is a result from Contra's own detector, not a universal measure of originality or quality. A page can avoid familiar AI design patterns and still be confusing, inaccessible, or wrong for a client's needs.

What we still do not know

The public leaderboard gives no way to inspect Gestalt's model weights, test it independently, or check how it behaves outside these briefs. The benchmark dataset is available by request, rather than as a public download.[2][3]

There are other limits. The task set is small, the briefs come from a specific marketplace context, and the evaluation centers on visual and functional outputs judged by designers. It does not, by itself, measure the full work of a design team: client interviews, changing requirements, handoff, maintenance, or long-term product outcomes.

These are limits of the evidence currently available, not proof that the benchmark is flawed. They define how far its results can travel.

The useful signal for designers

Even without access to Gestalt, the public benchmark offers one practical idea: evaluate design work with more than a single score. Contra combines blind preference, a detailed rubric, and a detector for recurring AI-style patterns. Each view answers a different question.

For a team comparing design tools, a simple checklist helps:

  • Test on briefs that resemble your real work, including less polished examples.
  • Hide model names when collecting human preferences.
  • Check usability and task completion as well as visual appeal.
  • Record the tools, time, and revisions needed to produce each result.
  • Treat benchmark scores as evidence about that test, not a promise about every project.

Gestalt's first-place result is a reason to watch the next update, not yet a reason to change a design workflow. The stronger evidence will come when the model is identified, made testable, and evaluated on more tasks by people outside the team that announced it.

What would you want to test first if Gestalt became publicly available?

Sources

[1] Frontier Design: Evaluating AI on Real World Design Tasks, Contra Labs
[2] Frontier Design Benchmark leaderboard, Contra Labs
[3] Frontier Design Benchmark dataset page, Contra Labs