Anonymous design champ Gestalt: Do stealth models keep their crown once unmasked?
A nameless model called Gestalt tops a new design leaderboard, ahead of human professionals and Claude Fable 5.2. Nobody knows who built it. We checked how much of their top spot earlier stealth models kept after their real names came out, and how solid the new record really is.

What this is about
Since November 19, a model with no known maker has topped a new leaderboard. On the Frontier Design Benchmark from Contra Labs and BenchFlow, professional designers judge blind which of two AI-built websites better fulfills a real client brief. At the top sits "Gestalt" with around 1,931 Elo points. Elo is a rating system borrowed from chess, where point gaps translate directly into win probabilities. Behind it come designs by human designers (1,739) and Claude Fable 5.2 (1,713). Contra describes Gestalt only as "a high-quality design model that is in stealth mode" and says it is working "with the team." Which team, Contra won't say.
Lists of which codename later turned out to be which model already exist. We're asking a different question: How much of their top spot did earlier stealth models keep once they appeared under their real names on the same leaderboard? And what does that mean for Gestalt?
What we know about Gestalt, and what we don't
The leaderboard is built on 33 tasks. 27 are real Contra jobs, rewritten for fictional clients; six come from startup partners. According to the methodology page, 35 designers cast 4,311 votes over five rounds. A Bradley-Terry model, a statistical method that turns pairwise comparisons into a ranking, converts these into Elo scores. The homepage, however, mentions "68 contributing designers." That figure probably includes the people who created the human reference designs, but it isn't spelled out anywhere.
The numbers for Gestalt itself don't line up either. According to Kingy.ai, Contra's November 18 report cites 378 wins from 423 votes and 1,928 points, while the leaderboard shows 420 votes and 1,931 points. Sulat AI reports that the 95 percent intervals of Gestalt and Fable don't overlap. The 35-criterion scoring rubric, which Contra says correlates with designer judgments at r = 0.94, doesn't put Gestalt in front, though: Fable 5.2 and Opus 5.5 score 84 out of 100, Gestalt 83. The report says the correlation covers 15 models, but the leaderboard lists 16. Whether Gestalt is included remains unclear.
The reaction on Hacker News was predictable. One user wrote "this is 100% an ad," another called it "black-hat marketing," since Gestalt doesn't show up on OpenRouter or anywhere else publicly. Contra co-founder Gajus Kuizinas replied only that it is "not a Contra model." There's no gallery of Gestalt designs. The theory that the model comes from Ilya Sutskever's company SSI was flagged as a rumor by the person who started it, according to SnowRock, and later called a probable joke. Others guess Anthropic, Zhipu or the startup Variant. There's no evidence for any of these. On top of that, there's a conflict of interest: Contra runs a marketplace for freelance designers and, by its own account, offers labs "high-quality data" to improve their models' design skills.
Our cross-check: anonymous vs. revealed
We collected the best-known cases where a model appeared on the same leaderboard before and after its reveal.
| Codename → real name | Leaderboard | Anonymous | After reveal |
|---|---|---|---|
| im-also-a-good-gpt2-chatbot → GPT-4o (May 2024) | LMSYS Chatbot Arena | approx. 1,309, rank 1 | 1,287 one week after listing, still rank 1 |
| summit → GPT-5 (Aug. 2025) | LMArena Text | Rank 1, "highest Arena score ever" | Days later just ahead of Gemini 2.5 Pro; today 1,435, rank 112 of 426 |
| nano-banana → Gemini 2.5 Flash Image (Aug. 2025) | LMArena Image Edit | 1,362, rank 1, 171-point lead | Rank 1 at launch; today rank 30 of 48 in "Multi Image Edit" |
| Llama-4-Maverick-03-26-Experimental (Apr. 2025) | LMArena Text | 1,417, rank 2 | Release version reportedly around rank 32 |
| maskingtape-alpha → GPT Image 2 (Apr. 2026) | LMArena | no score, removed after hours | 1,512, rank 1 (per trade media) |
The GPT-4o case is the best documented. On the prediction market Manifold, the question about gpt2-chatbot's Elo resolved at 1,287. One commenter noted that the earlier tweet had "greatly exaggerated" the final score. The drop was about 22 points, and the top spot held. For nano-banana, LMArena reported the "largest Elo lead in Arena history." On Artificial Analysis's rival leaderboard, Seedream 4.0 overtook it shortly afterwards, but that's a different leaderboard.
The only real crash was Llama 4, and the cause is clear: the Arena ran a variant that Meta itself described as "optimized for conversationality." LMArena said at the time that Meta should have made it clearer that this was a model tailored to human preferences. According to the paper "The Leaderboard Illusion" (Cohere et al.), Meta had tested 27 private variants before launch.
On OpenRouter, where most codenames surface (Quasar and Optimus Alpha → GPT-4.1, Horizon → GPT-5, Hunter and Healer Alpha → Xiaomi MiMo-V2), there's no quality ranking at all, only usage charts. Hunter Alpha sat near the top as long as it was free. After the reveal, MiMo-V2-Pro kept leading in April and May with around 4.7 to 4.9 trillion tokens per week, according to third-party analyses. That says something about price and integration, not quality. For Ox Alpha (GLM-5.3-Flash), an unverified score of around 80 percent on the DeepSWE coding benchmark circulated during the stealth phase. After the reveal, Z.ai itself reported 63.4 percent, and a community run came in at 58.4 percent.
Our take: The leaderboards themselves have rarely lied so far. Where the same model kept running under a new name, it held its spot at launch and its score dipped only moderately. Things got dangerous when the tested model wasn't the released one, or when rumors during the stealth phase were treated as measurements. In the long run, every winner loses its spot to successors. That's normal, not evidence of window dressing.
How many codenames were never resolved
The directory at gradually.ai counts 33 codenames since 2024. 22 are officially confirmed, six only attributed by third parties, and five never resolved: Cypher Alpha, Code Supernova, Giga Potato, Aurora Alpha and Crouton Alpha. Digital Applied's OpenRouter-only list has six without an official reveal among 13 historical entries. The directories partly contradict each other: gradually.ai lists Polaris Alpha as confirmed (GPT-5.1), Digital Applied only as suspected. Roughly speaking, one in six or seven codenames never gets a known maker, and another fifth never receives official confirmation.
What the platforms allow vendors to do
- LMArena (now Arena): According to its policy (as of September 1, 2026), vendors may test multiple variants anonymously. Anyone who releases early must confirm in writing that the test model is identical to the release, or it gets pulled from the leaderboard. Scores stay "preliminary" until enough fresh post-release votes come in. Results from anonymous tests that never ship go only privately to the vendor. LMArena considers the effect of variant selection small: around +11 Elo after 50 tests and 3,000 votes. The Cohere authors came up with 50 to 100 points in simulations.
- OpenRouter: Under its stealth terms of use, stealth models are time-limited and often free. Providers may use prompts for training. There's no quality leaderboard.
- Contra: We found no public rules for anonymous submissions, either on the leaderboard or in the methodology. Who submitted Gestalt, whether there were several private variants, and which tools, how many attempts and what costs were involved remain publicly unanswered. Kingy.ai notes that blind judging protects against brand bias, but doesn't prove that all entries were produced "under equal conditions."
How solid are 4,311 votes?
A sanity check speaks for the leaderboard. Under the Elo formula, a 218-point gap corresponds to an expected win rate of about 78 percent. That's exactly the 78 percent Kingy.ai reports for the head-to-head between Gestalt and Fable. Internal reproducibility is high: if you split the designers randomly into two halves, the rankings agree at 0.95, according to Contra.
The "better than humans" headline is shakier. It rests on 20 wins out of 28 matchups. By our calculation (Wilson interval, 95 percent), the true win rate lies somewhere between roughly 53 and 85 percent. That's barely above a coin flip, and far from a data base like nano-banana's, with over 2.5 million votes. Also, the votes come from a closed panel drawn from Contra's own network. Nothing can be verified from the outside.
The bigger picture
The history of stealth models doesn't make Gestalt more suspicious than any other. Testing anonymously and then revealing is routine. Chinese tech sites like smzdm openly describe a fixed pre-launch rhythm, for example with Kimi K3 ("Kivine") and Qwen3.8 Max ("Kaleb"). So far, genuine leaders have held their spot as long as the same model kept running.
But Gestalt differs from GPT-4o or nano-banana in three ways. There's no public access, the leaderboard is new and run by a company with its own business interests, and there's no rule guaranteeing that the test version and the release version are the same. Llama 4 showed how quickly a record evaporates when exactly that isn't the case.
Our take: The gap to Fable is too big to explain by chance or variant selection alone. But it only tells you something about finished websites from 33 briefs under unknown conditions. Take the record seriously once a vendor steps forward, the model is freely available and third parties confirm the results on a different leaderboard. Incidentally, those are the same conditions the Chinese commentators name. Until then, Gestalt is an interesting data point, not a proven breakthrough.
Sources
- Contra Labs: Frontier Design Benchmark (Rangliste) contralabs.com
- Contra Labs: Methodik contralabs.com
- Sulat AI: Contra's design benchmark puts stealth model Gestalt on top news.sulat.com
- Kingy.ai: Gestalt-Analyse kingy.ai
- TechBytes: Two stealth models just appeared on the leaderboard techbytes.app
- SnowRock: Mystery AI Beats Human Designers in Blind Test snowrock.com
- smzdm: 匿名模型Gestalt登顶设计榜 post.smzdm.com
- smzdm: 匿名刷榜是AI圈常规操作? post.smzdm.com
- gradually.ai: Stealth-Modelle-Verzeichnis gradually.ai
- Digital Applied: OpenRouter Stealth Model Census digitalapplied.com
- Manifold: What Elo will gpt2-chatbot have manifold.markets
- Simon Willison: gpt2-chatbot confirmed as OpenAI simonwillison.net
- Arena-Blog: Nano Banana arena.ai
- Beebom: Google reveals nano-banana beebom.com
- Arena: nano-banana Multi Image Edit arena.ai
- Arena: GPT-5 (High) Text Arena arena.ai
- OfficeChai: GPT-5 top model on LMArena officechai.com
- Habr: GPT-5 im LMArena-Ranking habr.com
- TechCrunch: Meta's vanilla Maverick ranks below rivals techcrunch.com
- The Register: Meta Llama 4 benchmark theregister.com
- Simon Willison: LMArena-Statement zu Llama 4 simonwillison.net
- arXiv: The Leaderboard Illusion arxiv.org
- Arena-Blog: Antwort auf The Leaderboard Illusion arena.ai
- Arena-Blog: Leaderboard Policy arena.ai
- OpenRouter: Stealth Program EULA openrouter.ai
- The Batch: Ox Alpha revealed as GLM-5.3-Flash deeplearning.ai
- AI for Automation: Xiaomi's secret 1T AI was #1 on OpenRouter aiforautomation.io
- MacGPU: OpenRouter May 2026 Rankings macgpu.com
- OfficeChai: Seedream 4.0 goes past Nano Banana officechai.com
- OfficeChai: ChatGPT Images 2.0 tops Arena officechai.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.