AI Model Comparison

Model profile & tracker

Gestalt: The Stealth AI Model Topping the Design Benchmark

Gestalt, an unnamed stealth model, leads Contra's Frontier Design Benchmark at 1,931 Elo. What is confirmed, who made it, and when you can use it.

Verified 2026-11-22

Updated 22 November 2026. This page tracks what is confirmed about Gestalt and gets rewritten whenever the leaderboard moves or the maker is identified.

Gestalt is an AI model with no named maker. On 18 November 2026 it appeared at the top of the Frontier Design Benchmark, a new leaderboard from Contra Labs and BenchFlow that scores models on real client design briefs, judged blind by professional designers. It has held first place since, ahead of every public model and ahead of the human reference designs.

The confirmed numbers

Rank Entrant Elo Win rate
1 Gestalt (stealth) 1,931 90%
2 Human designers (reference) 1,739 74%
3 Claude Fable 5.2 1,713 73%
4 Claude Opus 5.5 1,691 73%
5 GPT-6.1 Sol 1,668 70%

Source: Contra Labs launch report (18 November 2026) and the live leaderboard. The numbers move slightly as votes come in — the launch post showed 1,928 — so treat a few Elo as noise. The gap to the best public model, 218 points over Claude Fable 5.2, is not noise.

Against human reference designs specifically, designers chose Gestalt in 20 of 28 direct comparisons. Against individual models the pattern holds: about 78% of votes against Claude Fable 5.2, and a reported 100% against Kimi K3.

What the benchmark actually is

Thirty-three design briefs — 27 adapted from real jobs on Contra's marketplace, six from startup partners. Professional designers compare two outputs for the same brief without being told which model made which, or which side is human. Those blind choices become an Elo rating. As of 21 November the leaderboard reports 4,311 votes.

This makes it a preference benchmark, like LMArena, not a scored test. It measures whether professional designers would rather ship one output than another on commercial work — which is closer to real design jobs than pixel-similarity scores, but is still a preference signal, not proof of general design ability.

What is known and not known

Known: the scores above; the benchmark's structure (33 briefs, 4,311 votes, blind judging); that Gestalt has been on the leaderboard since 18 November.

Not known: who built it, what architecture it is, how it is invoked, what it costs, whether it is a single model or a multi-stage system, and whether it will ever be publicly available. There is no API, no technical report and no try-it page. As of 22 November it cannot be used by anyone outside the evaluation.

Who made it: the speculation so far

The name has fuelled guesses — "Gestalt" evokes gestalt psychology, which fits a model trained on how humans perceive whole designs. Two claims are worth recording because they came with names attached:

  • A post on X speculated Gestalt was Claude Fable 5.5 under a codename.
  • Siki Chen quoted that post and said Gestalt is not Fable 5.5 and not from Anthropic, OpenAI, Google, xAI or SSI — without publishing evidence.

Both are unverified. Only a lab claiming the model, or a leak of its harness or tokenizer, would settle it. Until then, any identity claim — including ones on this page — is speculation.

Should you wait for it

No. A lead on one preference benchmark, run by the benchmark's own publisher, on 33 briefs, is a reason to watch, not a reason to plan around. If you need a design-capable model today, the models ranked right behind it are public: Claude Fable 5.2, Claude Opus 5.5 and GPT-6.1 Sol — all in the comparison table and the model index. If Gestalt is released publicly, this page will carry its price, context window and availability the same week.

FAQ

Can I use Gestalt? No. There is no public access, no API and no waitlist as of 22 November 2026.

Is Gestalt better than Claude Fable 5.2? On this benchmark, clearly: a 218-point Elo lead and 78% of head-to-head votes. On other workloads there is no evidence either way, because no other results exist.

Is it better than human designers? On these 33 briefs, in blind votes, it beat the human reference set 20 times out of 28. That is a real result on a small sample. It is not a statement about design as a profession — see how leaderboards mislead.

When will we know who made it? Unknown. Stealth entries are usually either claimed at launch or never confirmed. This page updates when either happens.