TokenRa AI Model Catalog
Gestalt (stealth) · Frontier Design Benchmark

Gestalt (Stealth): the Anonymous Model at the Top of the Design Leaderboard

On 2026-11-18 an unnamed model called Gestalt appeared at the top of Contra Labs' Frontier Design Benchmark, ahead of every production model in the same run. No lab has claimed it, no model card exists, and there is nothing to call. This page records the confirmed numbers and the parts nobody has published.

1,928 EloFirst place on the Frontier Design Benchmark.
89% win rate378 of 423 blind designer votes.
0 endpointsNo public model ID, rate limit or price.

Last verified 2026-11-20 · Numbers from Contra Labs' benchmark write-up (published 2026-11-18).

Searching for gestalt psychology?

A search for the bare word gestalt is almost always about perception, gestalt principles or gestalt therapy — not about this model. This page is only about the stealth entry on the design leaderboard. See what Gestalt is not before you cite anything.

What Gestalt is

Gestalt is the code name of an anonymous model entered into the Frontier Design Benchmark, a design evaluation published by Contra Labs (the research arm of the Contra freelance platform) together with BenchFlow. "Stealth" is Contra Labs' own label in the leaderboard: the entry was measured, but the organisation behind it withheld its name, its identity details and its release plans.

In practice that means Gestalt is a result on a leaderboard, not a product. You cannot sign up for it, buy tokens for it, or point code at it. The only verified statements about it are the numbers Contra Labs published and the design outputs in their Golden Gallery.

What the benchmark actually measures

The Frontier Design Benchmark scores models on design work rather than on a prompt like “build a landing page for a coffee shop”. The setup, as published:

  • 33 tasks, of which 27 are rewritten from real design jobs posted on Contra; 6 come from startup partners and include the client's own brand assets (19 briefs start from a supplied logo, 17 ask for a mobile-first site).
  • 35 professional designers recruited on Contra cast 4,311 blind votes across five rounds, comparing two outputs on the same brief.
  • Votes are converted to Elo ratings with a Bradley-Terry model. Split-half repeatability is quoted at Spearman 0.95 averaged over 200 splits.
  • A separate 34-dimension critique rubric grades the same outputs and agrees with the Elo ranking at r = 0.94 — the benchmark's own check that the preference ranking is not arbitrary.
  • Human "golden designs" made by paid Contra designers are ranked in the same pool, so models are compared against professional work, not only each other.

The numbers Gestalt posted

MeasureGestalt (stealth)Claude Fable 5.2
Elo rating1,9281,714
Win rate on pairwise votes89% (378 of 423)Highest among production models in the run
Gap, with 95% intervals214 Elo points; the two intervals do not overlap
Head-to-head against human golden designsPicked 20 of 28 times—

For context on the rest of the field: six models sit in a tight top group between 1,638 and 1,714 Elo (three Claude models, two GPT models, Kimi K3.1 Pro), winning 66% to 73% of their votes. Nine others fall between 1,122 and 1,434, winning 12% to 41%. Gestalt is separated from that top group by more than 200 points.

Two caveats worth carrying into any argument you make with these numbers. First, Contra's live leaderboard and launch-day posts circulating on X display 1,931 and a 90% win rate; this page uses 1,928 and 378 of 423 because those are the figures the benchmark authors state in their write-up. Ratings move as votes accumulate, so any Elo is a snapshot. Second, the human golden designs win 74% of their own votes (425 of 574) while scoring only 73.6 on the rubric, sixth of sixteen — the benchmark's authors present this openly as evidence that the rubric measures avoidable mistakes, not taste.

Who made Gestalt?

Nobody has said. Contra Labs lists it as a stealth model and has published no lab name, no parameter count, no training details and no release date. The name itself is the only clue in circulation.

Threads about the leaderboard are already full of guesses, including the reading that "Gestalt" signals an ensemble of several models rather than one new model. Treat every one of those as a hypothesis: there is no model card, no weight release and no first-party statement to check against. We do not repeat the guesses as fact on this page, and you should not carry one into documentation or a pricing page of your own.

What you can verify first-hand is limited but real: Contra Labs' leaderboard, their research write-up, their methodology page, and the outputs in their Golden Gallery. Those are the primary sources. Screenshots, aggregator posts and "confirmed" identity claims in forums are not.

Is there a Gestalt API, model ID or price?

No — and this is the part most people searching for the model actually need. As of 2026-11-20 there is no public endpoint, no model identifier, no rate-limit table and no published price for Gestalt. Contra Labs ran the model inside their own benchmark harness, using OpenRouter as the model provider; that is a detail about how the evaluation was executed, not a listing you can call.

  • No API key or console exists for this name.
  • No model ID has been published, so any ID you see quoted in a tutorial is invented.
  • No pricing exists, per token or otherwise.
  • TokenRa does not route Gestalt. If that changes after an official release, this page will say so with a date.

Sceptical of any quoted key or rate

Be sceptical of any page that shows a Gestalt API key, a per-million-token price or a curl example for this model. Someone is either writing about a different product or filling the gap with invented detail — and the code names circulating in the same week make that mistake easy.

What Gestalt is not

The word pulls in a lot of unrelated traffic, and mixing these up is the fastest way to write something wrong.

TermWhat it refers toRelevance to this model
gestalt psychologyThe 1920s perceptual theory behind gestalt principles, gestalt therapy and gestalt coachingNone. This is what most searches for the bare word mean.
Gestalt AIA generic phrase used by unrelated tools, papers and consultancies describing systems that work as a wholeNone. Different projects use the same word independently.
Gestalt (stealth)The anonymous entry on Contra Labs' Frontier Design Benchmark, published 2026-11-18This page.

If you are writing about the design benchmark, say Gestalt (stealth) explicitly and name Contra Labs in the same sentence. If you are writing about perception, do not mention a model at all. Mixing the two produces a page that ranks for neither intent and misleads readers who arrive from either side.

What you can do with this today

Since there is nothing to call, the useful work is limited to three things — and none of them involve waiting for an API key.

1

Read the Golden Gallery rather than the score

The published outputs show what the model actually produced on real briefs. That is more informative than an Elo difference, because it tells you which design patterns it favours on a supplied brand asset versus a from-scratch page.

2

Compare against a model you can call

If your goal is design-generating output you ship, benchmark your own prompts against a model with published rates and a stable identifier, and score it on the same criteria Contra used: execution errors first, taste second.

3

Record today's numbers with their date

Elo ratings move as votes accumulate, and a stealth entry can disappear from a leaderboard without a single announcement. A figure quoted without its date is unusable within a couple of weeks.

What to watch next

Three signals would turn Gestalt from a leaderboard entry into a product. In each case, this page gets updated with a date rather than a rumour.

  • An identity announcement. A lab name, model card or weight release — the point at which speculation stops being speculation.
  • A published endpoint. An API console, model ID or rate table, whichever comes first.
  • A commercial release. Per-token pricing and rate limits under either the code name or an official name.

The opposite is just as likely: stealth entries on benchmark leaderboards sometimes never ship. If nothing is published in the coming weeks, the honest conclusion is that Gestalt was an evaluation entry, not a forthcoming model — and pages that promised an API will have to be corrected.

Need a model you can actually call today?

TokenRa routes paid model APIs with published per-token rates and stable identifiers. If you were looking for Gestalt because of the design scores, these are the closest things you can put behind an API key right now.

Gestalt (stealth) FAQ

What is Gestalt (stealth)?

An anonymous model entered into Contra Labs' Frontier Design Benchmark, published 2026-11-18. It ranks first with an Elo of 1,928 and wins 89% of its pairwise votes (378 of 423). No developer, model card, weights or API has been published under the name.

Is there a Gestalt API or model ID?

No. There is no public endpoint, model ID, rate limit or price. The model ran inside Contra Labs' benchmark harness with OpenRouter as the provider, which is not a public listing. TokenRa does not route Gestalt.

Who made Gestalt?

Not disclosed. Contra Labs lists it as stealth and has published no lab name, parameter count or training details. Names circulating in comment threads are guesses, not evidence.

Is Gestalt free to use?

There is nothing to use yet; it is not available at any price, free or paid. Any page quoting a free tier for this model is describing something else.

How does Gestalt compare with Claude Fable 5.2?

On this benchmark, 1,928 Elo versus 1,714 — a 214-point gap with non-overlapping 95% intervals. When designers saw a Gestalt site next to a human golden design, they picked Gestalt 20 times out of 28. This measures judged design output, not general reasoning or coding.

Is Gestalt the same as gestalt psychology or Gestalt AI?

No. Gestalt psychology is the 1920s perceptual theory behind gestalt principles and gestalt therapy, and it accounts for most searches on the bare word. Gestalt AI is also used independently by unrelated tools and papers. The stealth model is a separate, much newer thing.

Will Gestalt be released as a product?

Unknown. Contra Labs has announced no release date, and stealth entries sometimes never ship. This page will record an official name, endpoint or price with its date if one appears.

Where do these numbers come from?

Contra Labs' Frontier Design Benchmark write-up, published 2026-11-18: 33 tasks, 35 professional designers, 4,311 blind votes, Elo via Bradley-Terry. The live leaderboard and launch-day posts on X show 1,931 and a 90% win rate instead; ratings shift as votes accumulate.

Gestalt is a code name published by Contra Labs on its Frontier Design Benchmark, not by TokenRa. TokenRa is a routing gateway for model APIs and is not the developer or owner of the models listed here. Official names, benchmark figures and availability can change; provider and related names belong to their respective rights holders. Last updated 2026-11-20.