Stealth model · Frontier Design Benchmark
Gestalt (stealth)
A design model Contra Labs is previewing on its Frontier Design Benchmark. It tops the live board, ahead of every frontier model and of the human golden designs — and almost nothing else about it is public.
The short version
On the Contra Labs × BenchFlow Frontier Design Benchmark, read on November 22, 2026, Gestalt (stealth) ranks #1 of 17 entries with an Elo of 1,931 and a 90% win rate (95% CI 86–92) across 420 blind votes. The best model Contra actually ran, Claude Fable 5.2, sits at 1,713 — 218 Elo points lower. Contra Labs has not published who makes Gestalt, its architecture, its parameter count, a price, an API or a trial.
Read on 2026-11-22
Every figure on this page comes from the live leaderboard (header: "Updated Nov 2026 · 27 Contra jobs") opened on 2026-11-22, and from Contra Labs' November 18, 2026 research article. The board moves as votes come in, so treat these numbers as that day's snapshot. Where the article's earlier numbers differ, this page says so instead of blending them.
The live board
The top of the Contra Labs Frontier Design Benchmark, quoted row for row.
| Entry | Rank | Elo | Win rate | 95% CI | Votes |
|---|---|---|---|---|---|
| Gestalt (stealth) | #1 | 1931 | 90% | 86–92 | 420 |
| Human designers (Contra panel) | #2 | 1739 | 74% | 71–78 | 571 |
| Claude Fable 5.2 | #3 | 1713 | 73% | 69–76 | 683 |
| Claude Opus 5.5 | #4 | 1691 | 73% | 69–76 | 527 |
| GPT-6.1 Sol | #5 | 1668 | 70% | 66–74 | 549 |
| Claude Sonnet 5.5 | #6 | 1667 | 69% | 65–73 | 440 |
| Kimi K3.1 Pro | #7 | 1663 | 71% | 66–74 | 519 |
| GPT-6 Astra | #8 | 1638 | 66% | 61–70 | 449 |
Elo, win rate, 95% confidence interval and vote counts are copied from contralabs.com/frontierdesign/leaderboard as displayed on 2026-11-22. The board's own footnote puts the panel at 4,311 blind votes by 35 professional designers over 33 briefs, and its header counts 27 Contra jobs. Only the top eight rows are reproduced here.
Gestalt vs Claude Fable 5.2
The comparison this page exists for: where the stealth model sits relative to the strongest model Contra actually ran.
First, above the human reference
The board ranks Gestalt (stealth) #1 at 1,931 Elo. The second row is not a model: it is the panel of professional designers recruited from Contra, at 1,739. Contra Labs' article calls Gestalt the first model rated above the human golden designs.
218 points above Claude Fable 5.2
Claude Fable 5.2 — the highest-rated model Contra actually ran, per the November 18 article — is #3 at 1,713 Elo with a 73% win rate (95% CI 69–76) over 683 votes. Gestalt's win-rate interval (86–92) and Fable 5.2's (69–76) do not overlap.
The models behind it are tightly packed
Claude Opus 5.5 (1,691), GPT-6.1 Sol (1,668), Claude Sonnet 5.5 (1,667), Kimi K3.1 Pro (1,663) and GPT-6 Astra (1,638) sit inside a 53-point band. Gestalt's lead over that group is larger than the whole group's internal spread.
Design quality, not general capability
This is one benchmark of one skill: building a working website from a real client brief, judged blind by designers. Nothing on the board says anything about coding, reasoning or long-context work — and Gestalt has no published price to compare.
The gaps above are differences between the ratings the board displayed, not a significance claim. Contra Labs' own article reported a 214-point gap when the board read 1,928 for Gestalt and 1,714 for Fable 5.2.
How the ranking is produced
Enough of the method to judge the number.
33 real design briefs
Contra Labs built 33 tasks: 27 adapted from real design jobs posted on Contra, rewritten for fictional clients, plus 6 from startup partners. They cover multi-page marketing sites, landing pages, web apps with working flows and mobile apps.
35 professional designers, 4,311 votes
Every model attempts each task in an agent harness — BenchFlow's SDK, with OpenRouter as the model provider — and designers recruited from Contra vote on pairs of outputs.
Blind pairwise votes, turned into Elo
Reviewers see two outputs for the same brief side by side without knowing which model made which, and the sides are shuffled for every pair. The votes become per-model ratings through a Bradley-Terry model with Elo scores.
A design-critique rubric behind the ranking
Outputs are also graded on a curated rubric of design dimensions — type hierarchy, spacing, contrast, whether everything renders whole. Contra Labs reports the rubric's agreement with the designers' ranking at r = 0.94.
An AI-slop detector
A classifier trained on human-made and AI-made pages scores each output from 0 to 100 for how AI-like it looks. Contra Labs reports Gestalt's outputs as significantly less AI-looking than other models', close to the human golden designs.
Summarised from the November 18, 2026 research article and the benchmark's methodology page, both read on 2026-11-22. The article calls the rubric 34 dimensions and the methodology page calls it 35, so this page quotes no count.
What Contra Labs has not published
Everything below is unconfirmed, and this page does not fill it in.
Who makes it
No vendor, lab or team is named. The article says Contra Labs is "working with the team" on further evaluation without identifying it.
Architecture and size
No architecture, parameter count, training data or context window has been published. Anything you read elsewhere about Gestalt's internals is not from Contra Labs.
Release, API or trial
Contra Labs describes Gestalt as "currently in stealth" and offers no sign-up, waitlist, API endpoint or model ID. As of November 22, 2026 there is no public way to run it.
Price
No rate card exists, so the board's cost axis has no Gestalt point: the Elo comparison is quality-only.
This section goes out of date the moment Contra Labs or the model's team publishes any of it, which is why the page carries the date it was read.
The November 18, 2026 article
Contra Labs' research write-up predates today's board, and its numbers are slightly different.
Research article · 2026-11-18
The article that introduced the benchmark reports Gestalt (stealth) at 1,928 Elo, winning 89% of its votes (378 of 423), with Claude Fable 5.2 at 1,714 — "214 points apart", with non-overlapping 95% intervals. It also reports that, shown a Gestalt site beside a human golden design, designers picked Gestalt 20 times out of 28. The live board read on 2026-11-22 shows 1,931 and a 90% win rate over 420 votes for Gestalt, and 1,713 for Fable 5.2. Both sets are Contra Labs' own figures; they are different snapshots, not a correction.
Quoted from contralabs.com/research/frontier-design-benchmark (November 18, 2026), read on 2026-11-22. These figures may already be out of date — the leaderboard is the live source.
Sources
Opened on 2026-11-22; every number above is traceable to these three pages.
- Frontier Design leaderboard
Live Elo, win rate, 95% confidence interval and vote counts, header "Updated Nov 2026 · 27 Contra jobs".
- Frontier Design: Evaluating AI on Real World Design Tasks
Research article, November 18, 2026 — the "New Frontier Design Model Preview: Gestalt (Stealth)" section and the benchmark results.
- Methodology
Design taskset, model rollouts and human golden designs, the evaluation metrics and the Slop Detector.
How far this page can be trusted
This is one benchmark, run by one lab, on design output only. Gestalt's 420 votes are the smallest sample in the top eight, so its rating has the most room to move as voting continues; Fable 5.2's 683 votes make its 1,713 the better-sampled figure. The board is a snapshot — it read 1931 for Gestalt and 1713 for Claude Fable 5.2 on 2026-11-22. Nothing here is a review of the model, because there is no way to use it.
Live board, 2026-11-22: Gestalt 1931, human designers 1739, Claude Fable 5.2 1713, Claude Opus 5.5 1691, Claude Sonnet 5.5 1667. This site is not affiliated with Contra Labs, BenchFlow or Anthropic.
Compare it with the models you can actually use
Gestalt has no published price and no API. Claude Fable 5.2, Claude Opus 5.5 and GPT-6.2 do — start there if you need to ship something.