Best AI model for web design

No AI model wins web design by a margin a visitor would notice. On Design Arena's Website board at the time of writing (read November 10, 2026), Meta's Muse Wonder leads at 1367, and GPT-6 Astra, Kimi K3, Claude Opus 5.5, Kimi K3.1 Pro, Claude Sonnet 5.5 and Gemini 4 Argon sit within 18 points of each other. In that rating system the top of the group is expected to beat the bottom in about 53 votes out of 100.

So the useful answer runs against the question. The model stopped being the variable. What varies is how many design decisions were already made by the time the prompt arrived.

The short call, before the detail:

• Building a whole site in one pass? Gemini 4 Argon, whose output limit runs up to 1M tokens, or Claude Opus 5.5, the model Anthropic suggests starting with.
• Rebuilding a page from a screenshot? GPT-6 Astra, first on Design Arena's Image to HTML board.
• Running a long, reasoning-heavy build? Claude Fable 5.2, released November 9, or Claude Opus 5.5, first on the agentic board.
• Want open weights? Kimi K3.1 Pro or DeepSeek V4.1 Flash.
• Generating at volume on cheap tokens? Gemini 3.8 Flash, DeepSeek V4.1 Flash or Claude Haiku 5.5.
• Want the page to stop reading as generated? That is a brief problem, and Thanor sells the brief.

Each of these three was built from a prompt that had already decided what a model would otherwise fill with a default:

Every flagship clears the same bar#

Correct modern CSS is table stakes now. Every current flagship writes grid without being asked, sizes type with clamp(), respects prefers-reduced-motion, and produces a sticky header that does not jump on scroll. The floor of the category moved up, quietly, over the past year.

The bar itself moved, which is worth stating. A year ago a generated page needed hand repair before anybody saw it: a nav that broke at 768px, focus states missing from custom controls, transforms that ignored the accessibility query. Those faults are now the exception, so code quality is rarely the reason to prefer one model over another.

What none of them does is decide. Ask for "a modern SaaS landing page" and each one fills the same empty slots in roughly the same way, because the gap gets filled from the same distribution:

• Typeface: a neutral geometric sans, very often Inter.
• Neutral: a gray sitting almost exactly on the axis, #6B7280 and its immediate neighbors.
• Motion: opacity 0 to 1 plus translateY(20px), around 300ms, ease-out, applied to every section alike.
• Skeleton: hero, three feature cards, testimonial, call to action.

Read four outputs side by side and the family resemblance is not a flaw in any one model. It is what any model does with a slot nobody filled. The convergence is downstream of the instruction, which is the whole diagnosis of why generated sites rhyme.

So the useful comparison is not which model designs better. It is which six decisions you are willing to make before you press return.

The six decisions that change the page#

Six slots decide whether a page reads as art-directed or assembled. Every one of them is a value, and every one of them is yours.

Six lines of specification, and the same model that produced a template produces something with a point of view. The accent used once rather than six times is a decision a model almost never makes unprompted, because restraint is rarely anybody's default.

Two of the six rows are caps rather than values, and a cap is the thing a model does not write for itself:

• The accent appears on one element per section, never on body copy.
• The measure stops at 64ch however wide the grid is allowed to run.

The last column is the entire product category. A prompt that carries those values has made the choices; a prompt that says "clean, modern, premium" has made none, and the vocabulary gap between the two is measurable.

Where the models stand at the time of writing#

Design Arena ranks models by blind votes on sites built from the same prompt, which makes it the closest public measure of this article's question. How the board works, and how to read it covers the method. These rows are the Website board as it read on November 10, 2026.

Read the Elo column, not the rank. The leader is expected to beat Gemini 4 Argon in about 58 votes out of 100, and Claude Fable 5.2 had fewer than 600 battles a day after release, so its rating will still move. The time column spreads wider than the rating does: Argon averages under four minutes per site, Kimi K3.1 Pro almost twenty.

The order also changes with the job. On Image to HTML, where the input is a screenshot, GPT-6 Astra leads and Gemini 4 Argon is third. On the Overall Agentic board, where a coding agent builds across many files, Claude Opus 5.5 leads at 1459 and GPT-6.1 Sol is second. Arena's separate WebDev board puts Claude models in its top three, with Fable 5.2 listed first on an automated estimate rather than votes and Opus 5.5 second at 1818.

Two careful measurements that disagree on the order inside a group are telling you the group is close.

Which one to reach for, by job#

Tier and workflow matter more than brand once the brief is written. The current Claude models, GPT-6 Astra, Gemini 4 Argon, Kimi K3.1 Pro and DeepSeek V4.1 Flash all list a context window of about 1M tokens, so context alone no longer separates them.

The full site in one pass#

Gemini 4 Argon lists an output limit of up to 1M tokens, against 128K for the current Claude models and GPT-6 Astra, so a long single-file page with inline styles, scripts and copy fits in one response. Its access, prices and boards are in Gemini 4 Argon for web design. Claude Opus 5.5, released September 22, is the model Anthropic suggests starting with for most work, and Claude vs Gemini for web design compares the two lineups board by board.

One section, or a component#

Claude Sonnet 5.5, Claude Haiku 5.5, Gemini 3.8 Flash and DeepSeek V4.1 Flash are the fast tiers. The two Flash models average about 140 seconds per site on the Website board, so a hero can go round several times in the time a slow flagship takes once. A hero whose six values are already written down leaves the model little to decide, which is where a fast tier earns its place. DeepSeek V4 for web design explains which model the old V4 names now reach.

Copy, layout and long builds#

Claude Fable 5.2, released November 9, is Anthropic's model for demanding reasoning and long-horizon work, and the one to point at a page where the words and the composition have to agree. GPT-6 Astra holds the same place in OpenAI's lineup, with GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna below it.

Open weights#

Kimi K3.1 Pro, released October 23, publishes its weights on Hugging Face under its own license, and DeepSeek V4.1 Flash ships under the MIT License. For an agency that wants to host a model itself, start with Kimi K3 and K3.1 Pro for web design.

Claude Mythos 5.2 shares Fable 5.2's capabilities and is limited to organizations verified through Anthropic's programs, so it is news rather than a tool you can plan around. Claude Design, now a template inside Claude for designs and landing pages, exports standalone HTML or hands the work to Claude Code, which makes it a surface for a specification rather than a competitor to one. Claude Design prompts that don't give you the default look shows what to put in it.

What no model decides for you#

Refusals. Every real design is defined as much by what it will not do as by what it does, and a model rarely volunteers a refusal of its own.

A brief that names three of them moves the output further than the gap between the models above, which is a few votes in a hundred:

1. One accent, one place. Name the single element allowed to carry #D8492B and the page stops looking like a swatch demo.
2. No gradient behind text. A gradient arrives wherever depth was never decided, and it reads as filler because it is.
3. Nothing reveals twice. Motion on the first viewport only, and the rest of the page holds still.
4. No third typeface, no icon set. Both arrive to fill a gap and both add a vocabulary you did not choose.

Refusals have to be written per design rather than as general taste advice, because the useful ones contradict each other. A page whose argument is density refuses whitespace; a page whose argument is calm refuses the counter and the draw-on. Generic advice cannot hold both positions, which is why a library that states them per design produces work that diverges.

The second thing no model decides is the signature moment: one mechanism, on one section, described as physics rather than mood. A counter that eases with cubic-bezier(0.32, 0.72, 0, 1) over 240ms, a card that lifts 3px on hover, a mark that draws its own stroke once on load. Ask for "delightful micro-interactions" and you get four of them fighting.

Both of these carry one signature moment, written as a mechanism: the eyes follow the cursor.

Write the refusals down next to the six values and the brief is finished. That is roughly a page of text, and it is the same page whichever model you paste it into.

The verdict, named#

Pick the model you already have a subscription to. Claude Opus 5.5, Claude Fable 5.2, GPT-6 Astra, Gemini 4 Argon and Kimi K3.1 Pro sit close enough on the design boards that the choice is a workflow question more than a quality one. Then put your effort where it moves the result.

The ranking that matters runs the other way:

• The brief. What a leaderboard measures is a model's defaults, and a brief that names the values replaces them.
• The tier: speed, cost, and how much of a long page fits in one response.
• The maker: a few votes in a hundred at the top of the board, in an order that changes often.

That first line is the one worth keeping. The board ranks what a model does with an empty slot, and a specification leaves few empty slots, which puts the advantage in the document rather than the dropdown.

It also decides what to do with the next launch. A release raises the floor of the code and leaves the design where your brief left it, so the honest response to a new model id is to paste the same page of values into it and check that the values survived. That is a ten-minute job, and it is the whole upgrade process. The Opus 5.5 write-up applies the same check to one release.

If writing that document is the part you would rather skip, that is the job Thanor does: exact values per design, with the media the page uses, for personal and client work. Open Vesperion and copy its prompt; it is free with an account. Paste it into the model you already use and count the decisions it made before the model saw it. The rest of the library opens with a paid plan, and the Thanor MCP lets Claude, ChatGPT or Cursor pull a design's art direction without the clipboard.