Design Arena, explained: which model designs best

Design Arena is a public leaderboard at designarena.ai that ranks AI models by blind human votes on the things they build: websites, web apps, games, 3D scenes, slides and more. A voter types a prompt, four anonymous models build it at the same time, and the voter picks between them in pairs until all four are ranked.

At the time of writing (the boards were read on November 9, 2026), Meta's Muse Wonder leads the Website board at 1368 and Claude Opus 5.5 leads the overall agentic ranking at 1461. The number worth keeping is the gap. On the Website board, GPT-6 Astra, Kimi K3, Claude Opus 5.5, Kimi K3.1 Pro and Gemini 4 Argon sit within 18 points of each other, which in this rating system means the higher model is expected to win about 53 votes in 100.

The short call, before the detail:

• Want the model voters prefer for a one-page site? Any of the Website board's top ten. The order moves every week.
• Building a multi-file app with an agent? Read the Web Apps and Overall Agentic boards instead.
• Want a page that stops looking generated? That is decided by the brief more than the rank, and Thanor sells the brief.

Each of these three was built from a prompt that made every design decision before a model saw it:

How Design Arena works#

Every voting session follows the same tournament, according to Design Arena's methodology page:

1. A voter enters a prompt and picks a category, or the site picks one from the prompt.
2. Four models are drawn at random from the active pool, plus one backup. All four get the identical prompt and generate at the same time.
3. Two pairs are shown side by side with the model names hidden. Each choice is one vote.
4. Winners meet winners and losers meet losers, then a tie-breaker settles the middle. Five battles give five votes and a full first to fourth order.

The ratings come from a Bradley-Terry model rather than a running Elo update, although the site labels the column "Elo" and starts every model at 1200. The model estimates each entrant's strength from all its pairwise results and converts it with Rating = 400 × log10(strength). Every vote counts the same, with no editorial weighting.

Three details change how the numbers should be read:

• Names stay hidden throughout the vote, so a brand cannot carry a rating.
• Every row carries a margin of error. In the Website top fifteen it runs from ±0.5% on the most-voted entries to ±2.2% on the newest.
• A model is marked preliminary until it reaches roughly 200 comparisons, and models under 50 are left off the main charts.

The Builder category works differently. There the same prompt goes one-shot to app builders such as Lovable, v0 and Bolt, and voters compare the finished apps. Design Arena calls it "an initial, best-effort procedure", which is a fair warning to read it as a snapshot.

The boards that matter for web design#

Design Arena runs dozens of boards across code, office documents, images, video and audio. Eight of them bear directly on building a site, and each one judges a different job.

For a landing page or a marketing site, the Website board is the closest match: one prompt, one file, judged as a rendered page. Most Thanor prompts, Ardente's included, ask for the same output, one standalone HTML file with one style block and one script, which is why this article leans on that board.

The agentic boards answer a different question. There each model runs inside a harness with file tools, a shell, web search, image generation and a deploy step, according to Design Arena's harness documentation, and the vote judges a project built across many files and tool calls.

GPT-6 Astra leads four of the eight boards, but never by much on the ones closest to a website. The Builder board is a pool of its own, so its numbers do not compare with the model boards: Replit's Design Mode leads, Lovable sits fourth at 1243, Figma Make tenth at 1211 and v0 thirteenth at 1190. If you are choosing between those tools, the Bolt and v0 comparison covers what each is good for, and the wider field of Bolt alternatives adds Framer, Webflow and Cursor.

Who leads the Website board right now#

These are the top ten on the Website board as the page showed them on November 9, 2026, at 8:05 PM UTC. The board lists more than 180 models in all.

Read the Elo column, not the rank column. From third to tenth is 19 points. Under the board's own formula, the leader is expected to beat the tenth model in about 58 votes out of 100, and Kimi K3 is expected to beat Gemini 4 Argon in about 52. Those are real preferences, measured across thousands of votes, and they are also small ones.

The time column is easy to skip and worth reading. Muse Ember averages 33 seconds per site; Kimi K3.1 Pro averages almost 20 minutes. Both sit in the top eight. If you iterate on a hero ten times in an afternoon, that column matters more than six Elo points.

Further down, Claude Fable 5.1 sits at 1306 and DeepSeek V4.1 Flash at 1305. Gemini 4 Argon's 1315 is 85 points above Gemini 3.1 Pro Preview at 1230, the Pro model the Gemini 3 write-up covered. Claude Fable 5.2, which Anthropic announced on November 9, has no rating yet. Each of those models has its own write-up in this series: Kimi K3 and K3.1 Pro, Gemini 4 Argon and DeepSeek V4.

Design Arena is not WebDev Arena#

WebDev Arena is a separate leaderboard. It is run by Arena, the project that was called LMArena (lmarena.ai now redirects to arena.ai), and its web board is called Code Arena: WebDev. Different voters, different prompts and a different scale: its scores sit around 1700 to 1800.

On November 8, 2026, the top of its overall WebDev board read:

1. Claude Opus 5.5 (Max), 1818
2. Claude Sonnet 5.5 (xHigh), 1774
3. Gemini 4 Carbon (High), 1768, marked preliminary
4. GPT-6 Astra (Max), 1767
5. Claude Fable 5.1 (Max), 1745

Further down, Kimi K3.1 Pro (Max) is tenth at 1688 and Gemini 4 Argon (High) eleventh at 1683. Earlier in the year the same site put Kimi K3 first on its Frontend board at launch, with Arena reporting a 76% pairwise win rate against 63% for Claude Fable 5.

Two honest boards, two different orders. Muse Wonder leads Design Arena's Website board and its full variant is seventh on WebDev; Claude Opus 5.5 is seventh on one and first on the other. What both boards agree on is the set: largely the same models from Anthropic, OpenAI, Google, Meta and Moonshot fill both top tiers. When two careful measurements cannot agree on the order inside a group, the group is close.

So compare positions, never numbers, across the two sites. A 1325 on Design Arena and an 1818 on WebDev do not describe a gap between them; they describe two scales.

What a blind vote measures#

A vote is a preference between two rendered pages built from the same prompt. Whatever the prompt left open, each model decided for itself: the typeface, the neutral, the accent, the motion, the order of the sections. So the board ranks defaults, the choices a model makes when nobody else made them.

That is a useful thing to rank, and it is not the thing you ship. The defaults are also where generated pages start to rhyme. Left blank, every flagship reaches for a neutral sans, a gray near the axis and a fade-and-slide at about 300ms, which is the whole diagnosis of why AI sites look alike. A model can win the vote with better defaults and still hand you a page that looks like everyone else's.

One board hints at how much the input moves the order. When the input is a screenshot instead of a sentence, on the Image to HTML board, the ranking reshuffles:

That table does not prove the input causes the change, since the tasks differ too. It does show that a model's rank depends on what it was handed, and a brief full of decided values is closer to a screenshot than to a one-line prompt.

How to use the board for your own site#

The leaderboard is a good shortlist and a poor decision. Four steps turn it into something you can act on.

Match the board to the job#

A single-page site or a hero belongs to the Website board. A React app with routes belongs to Web Apps. Rebuilding a reference belongs to Image to HTML, and a 3D product scene to 3D Design. A model can sit tenth on one and third on another, as Gemini 4 Argon does.

Treat 20 points as a tie#

Inside about 20 Elo points the expected split is closer than 53 to 47. Check the battles column too: a model with 249 battles and a model with 31,141 are not measured equally well, and the preliminary label is there for a reason.

Weigh time and price#

The site shows average generation time per row and charts rating against speed and against price per million tokens. A model that takes a minute lets you run ten variants while a five-minute model finishes two.

Then write the brief#

Pick from the top ten on workflow, price and speed. Then put the effort into the document you will paste, because that is the part of the result the rank does not cover. The pillar on choosing a model lists the six decisions the brief has to make.

Write down the decisions the vote rewards#

A voter judges a rendered page in seconds, so what wins is what a visitor sees first: the hero, its picture, its type and its first movement. A brief that decides those things as values takes the decision out of the model's defaults. Six lines carry most of it:

palette   ground #0B1220, text #F2EFE6, accent #FF5B14 on one element per section
type      display 96px at 0.92 line height, body 17px at 1.6, ratio 1.25, two weights
motion    reveal cubic-bezier(0.16, 1, 0.3, 1) over 700ms, first viewport only
hero      one film, scrubbed by the scroll over 3 screens, poster shown first
order     hero, details, specification, workshop, contact, footer
refuse    no gradient behind text, no gray on the neutral axis, no icon set

That fragment is illustrative, written for this article. A full art direction goes much further: every color with its role, contrast measured under the glyphs, the media files the page uses, and what happens with scripts off or reduced motion. These two show how far a specification can push a single screen:

Rookery is the clearest case of a refusal doing the work. Its brief allows exactly two colors, #E8EEF1 and #0E1A2B, with no accent and no gradient, and the film itself was mapped to those two values line by line. No default produces that palette, whatever the model's rank.

The verdict on Design Arena#

Design Arena is the best public measure of which model voters prefer for a generated page, and it is worth checking before you commit to a subscription. Read it as a shortlist.

What it says, at the time of writing:

• For a one-page site, Muse Wonder leads, and nine more models from six makers sit within 53 points of it.
• For agent-built projects, Claude Opus 5.5 leads the overall agentic board by 36 points over GPT-6.1 Sol.
• For rebuilding a reference, GPT-6 Astra leads and Gemini 4 Argon climbs to third.

What it does not say is how your page will look, because your page is built from your prompt, not the arena's. A rank is earned on whatever a voter typed that day; your page is built from the values you decided. The closer your brief comes to a finished specification, the less the model's defaults are left to judge. For a direct comparison of two of the makers on those boards, read Claude vs Gemini for web design or GPT-6 Astra vs Claude, run on one brief.

Run the arena test on a real brief#

The useful experiment is the one the arena cannot run for you: the same complete specification in two models from the top ten. If the values survive in both, you have learned that the model is a workflow choice and the brief is the design.

Open Vesperion and copy its prompt; it is free with an account. Paste it into the model at the top of the board you care about and into the one in tenth place, then compare the two pages side by side, exactly as a voter would. If you work in Claude, ChatGPT or Cursor, the Thanor MCP pulls the same art direction into the session without the clipboard. The rest of the library works the same way, one finished brief per design.