AI design leaderboard: Gestalt, Claude Fable 5.2 and the Frontier Design Benchmark
Short answer: Gestalt, an anonymous model with no named maker, leads Contra Labs’ Frontier Design Benchmark at Elo 1931 and a 90% win rate, ahead of the human-designer baseline (1739) and Claude Fable 5.2 (1713). The benchmark launched on 18 November 2026 and is built on 33 real client briefs. No lab has claimed Gestalt, no release date is published, and it has no public API price — so the question worth tracking is what the board says next, not when you can buy it.
Board snapshot: 21 November 2026, the leaderboard as published at launch. Contra updates it in rounds; this page keeps the date it was read.
The board, as it stands
Seventeen entries, one of them a human baseline rather than a model. Win rate is the share of blind pairwise votes an entry wins, with a 95% confidence interval; votes is how many comparisons it appeared in. Prices are the list rates this site checks per model, shown only where a provider publishes them.
| # | Entry | Maker | Elo | Win rate (95% CI) | Votes | List price /1M (in / out) |
|---|---|---|---|---|---|---|
| 1 | Gestalt | Undisclosed (stealth) | 1931 | 90% (86–92) | 420 | — |
| 2 | Human designers | Contra network baseline | 1739 | 74% (71–78) | 571 | — |
| 3 | Claude Fable 5.2 | Anthropic | 1713 | 73% (69–76) | 683 | $8 / $40 |
| 4 | Claude Opus 5.5 | Anthropic | 1691 | 73% (69–76) | 527 | $4 / $20 |
| 5 | GPT-6.1 Sol | OpenAI | 1668 | 70% (66–74) | 549 | $2 / $10 |
| 6 | Claude Sonnet 5.5 | Anthropic | 1667 | 69% (65–73) | 440 | $2 / $10 |
| 7 | Kimi K3.1 Pro | Moonshot AI | 1663 | 71% (66–74) | 519 | — |
| 8 | GPT-6 Astra | OpenAI | 1638 | 66% (61–70) | 449 | $10 / $50 |
| 9 | Qwen3.8 Max | Alibaba | 1434 | 39% (35–44) | 428 | — |
| 10 | Kimi K3 | Moonshot AI | 1415 | 40% (35–44) | 425 | — |
| 11 | GLM-5.3 Prime | Z.ai | 1403 | 41% (37–45) | 464 | — |
| 12 | Gemini 4 Carbon | 1366 | 36% (32–40) | 546 | $2 / $10 | |
| 13 | Muse Spark 1.3 | Meta | 1335 | 31% (27–35) | 503 | — |
| 14 | Muse Wonder | Meta | 1249 | 22% (18–25) | 486 | — |
| 15 | Grok 4.7 | xAI | 1247 | 23% (19–26) | 559 | $2 / $6 |
| 16 | Gemini 4 Argon | 1221 | 20% (17–24) | 547 | $2 / $10 | |
| 17 | DeepSeek V4 Pro | DeepSeek | 1121 | 12% (9–15) | 506 | $1.32 / $3.96 |
Read the last three columns together. Gestalt's 90% rests on 420 votes, the smallest sample on the board; Claude Fable 5.2's 73% rests on 683. A 218-point Elo gap is large, but the two numbers are not measured on equally firm ground yet, and the interval on Gestalt (86–92) is the widest of the leaders.
What the Frontier Design Benchmark measures
Contra Labs is the research arm of Contra, a marketplace where companies hire designers. The benchmark, built with BenchFlow, uses 33 briefs: 27 rewritten from real Contra jobs with the client renamed, and 6 supplied by startup partners with their own brand files. BenchFlow runs an agent on each brief in a sandbox and records the output, its cost and the full trajectory; the models were served through OpenRouter.
Designers from Contra's network then voted blind. They saw two outputs for the same brief side by side, with the model names hidden and the sides shuffled, and picked the better one. 4,311 votes over five rounds became a Bradley–Terry rating, published as Elo. To test stability, Contra split the panel in half and re-ranked the models with each half: the two rankings agree at Spearman 0.95 across 200 splits.
Two more scores sit beside Elo. A 35-item Design Critique Rubric grades execution — font count, spacing consistency, contrast, hierarchy — and tracks the designers' ranking at r = 0.94. A Slop Detector scores how much an output looks AI-made; lower is better. That is a more useful pair than a single number, because it separates “designers preferred it” from “it is well built”.
Contra's own pages do not fully agree on the sample: the methodology says 33 briefs and 35 designers, the leaderboard header says 27 Contra jobs, and the project page counts 68 contributing designers. The Elo votes are the 4,311 cast by the 35-designer panel; the other figures describe the wider project. When you quote the rating, quote the vote count with it.
Who is Contra Labs?
The research arm of Contra, the freelance marketplace where the briefs came from. Contra Labs runs the benchmark with BenchFlow and has published the Human Creativity Benchmark and a yearly Creative Intelligence Report; the design board is its public model leaderboard. Nothing ties it to Gestalt: it presents itself as the evaluator, and it does not sell any of the models it ranks.
Why a 90% win rate is not a 90% chance of a better design
- Sample size. Gestalt appeared in 420 comparisons, Claude Fable 5.2 in 683. The confidence intervals overlap with nothing at the top, but a new entrant with few votes is the entry most likely to move after the next round.
- Preference is not craft. On Contra's rubric Gestalt scores 83, behind Claude Fable 5.2 and Claude Opus 5.5 at 84 each, while the human golden set scores 74. Gestalt wins the taste vote and trails slightly on execution checks. Taste is also the part that varies most between briefs.
- The human comparison is close. In direct comparisons against human golden designs, designers chose Gestalt 20 times out of 28. A clear win, not a shutout — and the human set is a bar the benchmark built, not a team competing for a prize.
- Elo is relative to the pool. 1931 means “beats these seventeen entries”, not “1931 out of 2000”. Design Arena's leader sits near 1360 with a different pool and anchor, so the two Elo numbers cannot be compared with each other.
Gestalt vs Claude Fable 5.2
| Gestalt | Claude Fable 5.2 | |
|---|---|---|
| Maker | Undisclosed | Anthropic |
| Status | Stealth, no product page | On sale since 9 November 2026 |
| Frontier Elo | 1931 | 1713 |
| Win rate | 90% (86–92) | 73% (69–76) |
| Rubric score | 83 | 84 |
| API price /1M | None published | $8 in / $40 out |
| Context / output | Unknown | 1M tokens / 128K output |
| Evidence | Benchmark entry only | Model card and pricing docs |
Fable 5.2 is the highest-ranked model you can actually buy: 1713 Elo, 73%, $8 in and $40 out per 1M tokens, a 1M-token context and 128K output. Claude Opus 5.5 (1691, $4/$20) and GPT-6.1 Sol (1668, $2/$10) follow. At 10K input and 5K output tokens per generation — a rough single-page build — one generation costs $0.28 on Fable 5.2, $0.14 on Opus 5.5 and $0.07 on GPT-6.1 Sol at list prices. Gestalt has no price, so it has no cost per generation; any figure you see for it is invented.
The honest read: Gestalt leads on designer preference by a margin the panel can measure, but it is not a product. Fable 5.2 trails by 218 Elo and is available today with a contract, a data policy and a price. Ship on Fable this month; keep the board open for the launch.
Who made Gestalt?
Nobody has said. Contra Labs labels the entry “Stealth” and publishes no team, architecture, training data or hosting details. The model was served through OpenRouter for the benchmark, which means a provider had API access provisioned — it does not say who the provider is.
OpenRouter's public model list (checked 21 November 2026) does not include Gestalt, and no endpoint documentation, price or terms have been published. Chinese tech press describes a stealth endpoint available for calls; a listing that is not documented is not a product you can plan around. Treat any claim that names a lab as a guess until one of these appears: a model card, a pricing page, a statement from the lab, or Contra naming the provider.
The name is a design term — the gestalt, the whole that reads as more than its parts — which fits a benchmark about composition. It is not evidence of who built the model.
How to use it today
- You cannot call it as an API. There is no price, no rate limit, no data policy. If a stealth endpoint appears on a marketplace, it is an evaluation build: no SLA and unknown logging. Do not send client work through it.
- You can see its output. Contra publishes the briefs and the outputs side by side in the golden gallery, and the blind vote shows pairs with the names hidden. That is the closest thing to a trial.
- You can use the runner-up. Claude Fable 5.2, Claude Opus 5.5 and GPT-6.1 Sol are the top three purchasable models on the board, and the Claude cost calculator prices a design pipeline on them.
When will it launch?
No date, no price, no terms. The benchmark went live on 18 November 2026 and the coverage that made “Gestalt” a search term arrived on 20–21 November. Contra has said nothing about a release. Anyone giving you a month is extrapolating from other stealth launches, not from this one.
Stealth models usually ship under a new name. When that happens, “Gestalt” stops being the thing to search for; the benchmark, its rubric and this page do not. That is why this is built as a tracker of the design-capability board rather than a page about one codename.
What this page will change, and when
- Board snapshot. Re-read while the ranking is moving; the date at the top is the read date, not a publish date.
- Released name. Added to the codename table below when a lab claims the model, with the evidence link.
- Price and cost per generation. Added only when a provider publishes a rate; until then the column stays empty rather than estimated.
- Rubric and slop scores. Added when Contra prints them in text; the values here come from its published charts.
- Retirement. If the model is withdrawn instead of launched, the row is marked retired and kept for the record.
Codename watch
| Codename | Status | First seen | Claimed by | Released as | Last checked |
|---|---|---|---|---|---|
| Gestalt | Stealth, no API price | 18 Nov 2026 (benchmark launch) | Undisclosed | — | 21 Nov 2026 |
This row is the part that survives a rename: the benchmark entry, the rubric score and the vote count stay attached to the model even after it gets a product name.
Other design leaderboards, and why they disagree
Design Arena ranks single-turn HTML web design on crowd votes; in June 2026 GLM-5.2 took first place from Claude Fable 5 there, at around 1360 Elo. LMArena's WebDev Arena runs pairwise web-development battles with a crowd. Frontier Design uses 35 professional designers, 33 multi-page briefs and a 35-item rubric.
The pools differ, so the winners differ. A crowd rewards what looks impressive at a glance; a designer panel punishes inconsistent spacing and weak hierarchy. Neither is wrong — they answer different questions. For a landing page that has to convert, the crowd is closer to your user. For a multi-page brand site, the panel is closer to your client.
What this means for cost
Design quality is only half of a build decision; the other half is what one attempt costs and how many attempts it takes. A model that is 218 Elo behind but available at a published rate can still be the cheaper route to a finished page, because the expensive part of design work is the iteration, not the first generation. Price the iteration on your own token counts with the cost calculator before you switch.
Sources
- Contra Labs, Frontier Design Benchmark leaderboard — Elo, win rates, intervals, votes; read 21 Nov 2026.
- Contra Labs, methodology — 33 briefs, BenchFlow rollouts, 4,311 votes, rubric and slop detector; read 21 Nov 2026.
- Anthropic, Claude Fable 5.2 model page — release date, context and output limits.
- IT之家 / C114 report — rubric scores (Gestalt 83, Fable 5.2 and Opus 5.5 84, human 74) and the 20-of-28 head-to-head, from Contra's published charts.
- OpenRouter model list — checked 21 Nov 2026; no Gestalt entry.
- TechRadar on Design Arena — GLM-5.2 taking first place from Claude Fable 5, June 2026.
Questions
Is Gestalt a real model or a benchmark entry?
Both, in a narrow sense: it is a model that produced designs for the benchmark and was voted on blind, but it has no product page, no public price and no named owner. Treat it as a measured model, not as something you can use today.
Who made Gestalt?
Unknown. Contra Labs labels it Stealth and publishes no team, architecture or hosting details. The model was served through OpenRouter for the benchmark, but OpenRouter’s public model list did not include it when checked on 21 November 2026. A lab name in a blog post is a guess until a model card, pricing page or lab statement exists.
Is Gestalt better than Claude Fable 5.2?
On designer preference, yes: 1931 Elo against 1713, and a 90% win rate against 73%. On execution, not clearly: Contra’s rubric gives Gestalt 83 and Fable 5.2 84. Fable 5.2 is also the one you can buy, at $8 in and $40 out per 1M tokens.
What is Contra Labs?
The research arm of Contra, a marketplace where companies hire designers. It publishes the Frontier Design Benchmark with BenchFlow, and previously the Human Creativity Benchmark and a Creative Intelligence Report.
Did two stealth models top the board?
No. One model, Gestalt, is first; the second bar is a human-designer baseline built from Contra’s network, not a second anonymous model. Some coverage misread the two bars as two models.
How often is this page updated?
The board is re-read while the ranking moves, and the read date is printed at the top. The codename table gets a new row only when there is a source: a release, a price, or a claim from a lab.
Next step → Price a design generation on the model you can actually buy
Price data last checked 2026-11-18 against official sources · how we verify
Price-change alerts
One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.