Best LLM for UI/UX Design: 59 Generations, Measured

Key takeaway: GenDesigns tested six LLMs on ten identical mobile UI prompts in August 2026, producing 59 successful generations. Claude Opus 5 built the most interface — a median 201 HTML elements and 24 inline SVG icons per screen — while GPT-5.6 Luna Pro returned 109 elements in 25 seconds with the tightest latency spread of any model. Visual quality was not scored in this run.

Almost every page answering "best LLM for UI/UX design" is somebody's opinion with a table attached. No prompts published, no numbers, no date on the claim. That is a problem when the models being ranked ship new versions every few months.

So here is the alternative: everything below is either a measurement we ran, dated and reproducible, or a measurement we ran five months earlier that we are explicitly labelling as belonging to different models. Where we did not measure something, we say so and stop.

That distinction matters more than usual on this question, because "best for UI/UX" splits into two things that need different evidence:

• Does the model reliably produce a working, complete, reasonably dense interface? Machine-checkable. We measured it on 8 August 2026.
• Does the output look good and is the information architecture right? Not machine-checkable. It needs human judgement, and our human panel last convened in March 2026 against a completely different set of models.

Both halves are below. Neither is dressed up as the other.

The short answer

August 2026 run. 10 prompts x 6 models x 1 run = 60 generations, 59 successful. Llama 4 Maverick failed one generation outright.

If you want a one-line pick from this data alone:

• Most interface per call, most consistent icon output: Claude Opus 5.
• Best speed-to-substance ratio, most predictable response time: GPT-5.6 Luna Pro.
• Most markup per second: Gemini 3.6 Flash.
• Cheapest tokens per element: Grok 4.5.
• Not viable for finished screens: Llama 4 Maverick.

None of those bullets is a claim about which output looks better. Read on for why that distinction is the whole point of this page.

What we measured in August 2026

The harness is scripts/llm-ui-benchmark.ts in our repo. Same system prompt for every model, no per-model tuning:

You generate mobile app UI as a single self-contained HTML document using Tailwind CSS via CDN. Target a 390px-wide mobile viewport. Output only the HTML document. No explanation, no markdown fences.

Ten prompts, spanning simple to complex:

Five things came out of it that actually change how you pick a model.

1. Format compliance is no longer a differentiator

Every model returned a complete HTML document, every model loaded Tailwind from CDN, every model declared the viewport, and no model leaked a markdown fence. Six models, 59 generations, zero format failures.

That is a boring table and an important result. "Follows instructions" was a real tiebreaker a year ago. In this cohort it separates nobody. If a comparison page is still ranking models on whether they wrap output in prose, it is ranking a solved problem.

The one reliability finding is a hard one: Llama 4 Maverick failed p3 entirely with an Invalid JSON response from the provider. We are reporting it rather than retrying quietly. A 1-in-10 hard failure is a production fact, not a rounding error.

2. The spread is speed versus substance, and it is roughly 6x

Claude Opus 5 produces about six times more markup than Llama 4 Maverick and takes about six times longer. Everything else sits between.

The number worth internalising is not median latency but elements per second:

Llama 4 Maverick has the fastest median response time in the run and the slowest rate of building interface. Its 15.8-second median looks like a win until you divide by what arrives. That is the trap in every latency-only comparison: a fast empty answer is not a fast answer.

3. Icon output is where models diverge most for UI work

This is the metric we added specifically because it is a UI question rather than a general-capability question. The harness counts inline <svg> elements in each document. With no external asset host available, an inline SVG is essentially the only way a model can draw a real vector icon.

Two things stand out.

Claude Opus 5 draws roughly 17x more vector icons than Llama 4 Maverick over the same ten prompts (223 versus 13). For mobile UI that is not a cosmetic difference. Tab bars, toggles, chevrons, rating stars, engagement actions, back arrows — these are icons or they are nothing.

The zero-SVG column is the more useful one. Claude Opus 5 and GPT-5.6 Luna Pro never shipped an icon-free screen. Gemini 3.6 Flash shipped four, including a 201-element food delivery listing and a 155-element settings screen with no vector icon anywhere in the document. Grok 4.5's single miss was p4, the social feed — 150 elements including engagement actions, zero inline SVG. Whatever stands in for "like" and "comment" in that output is not a vector in the document. It might be a text glyph, an emoji, or a CSS shape; it is not an icon you can restyle.

If your workflow is "generate, then hand to a developer," a model that is bimodal on icons is worse than a model that is uniformly modest. Consistency beats peak here.

4. Every model scales with prompt complexity — from a different floor

We compared each model's element count on the two most component-dense prompts (p5 food delivery, p7 SaaS dashboard) against the simplest one (p9, an onboarding step with an illustration area, headline, body, and one button):

The ratios are remarkably tight — 2.1x to 2.7x across models whose absolute output on these prompts differs by roughly 5x. No model in this cohort ignores complexity. They all build proportionally more for a harder prompt. What differs is the floor they build from.

That reframes the Llama result. It is not failing to notice that a food delivery listing is harder than an onboarding slide; it responds to that just like the others. It responds at a resolution that does not produce a usable screen. Do the arithmetic on its 45-element p5 output: subtract the document scaffolding, and you have room for a search field, a few filter chips, and roughly four cards. That is a functional sketch. Claude's 295 elements on the same prompt is a different category of artifact.

It also puts a ceiling on what element count can tell you. High counts are not automatically better — 295 elements could be a well-built screen or an over-nested one. We did not open the documents and score them, so we are not going to tell you which.

5. Predictability is a real axis, and GPT-5.6 Luna Pro owns it

GPT-5.6 Luna Pro's ten generations all landed inside a 9-second window. DeepSeek V4 Pro's slowest single generation took 223.8 seconds — the longest in the entire run — for the e-commerce product detail screen. Llama 4 Maverick returned one screen in 4.1 seconds and another in 87.0 seconds.

If you are building a product where a user watches a spinner, that variance is a design constraint, not a footnote. A model with a 1.5x spread lets you write an honest loading state. A model with a 21x spread does not.

One more billing-adjacent number, since output tokens are what you actually pay for:

GPT-5.6 Luna Pro burns roughly twice the tokens per element of Claude Opus 5 while producing about half the markup. Token count is a cost number, not a measure of what you received — and the two come apart badly at the extremes.

What the March 2026 panel measured — and why those scores stay dated

The August run says nothing about whether the output is any good to look at. For that we have a separate study, and it is five months old and about different models.

In March 2026, a panel of three people — a designer, a developer, and a product manager — independently scored 150 generations across five weighted criteria: visual quality (25%), layout logic (25%), code quality (20%), component accuracy (15%), consistency (15%). Full method and per-model breakdowns are in the LLM UI generation benchmark.

Those scores belong to the March cohort:

The useful thing that panel established is not a leaderboard. It is a set of axes that turned out to be real: layout logic and visual polish are separable, and the same model is rarely best at both. Claude 4 Opus led layout logic at 9.0 and lost visual quality to GPT-4o. GPT-4o produced the prettier multi-step checkout and put the order summary in the wrong place. Gemini 2.5 Pro was the only model to nail Material Design 3 and the weakest at glassmorphism.

That finding — ask which axis you need, not which model is "best" — survives a model refresh. The numbers do not.

Why we will not carry those scores forward

Look at what actually changed between the two runs:

Not one model appears in both rounds. The Gemini row is the clearest case: reading Gemini 2.5 Pro's March visual-quality score as a prediction for Gemini 3.6 Flash would compare a frontier-tier model to a speed-tier one across a major version change. That is not a stale number, it is a wrong one.

So the panel scores stay on their own page, labelled March, for the models they were measured on. When we re-convene the panel it will use the same ten prompts, the same five weighted criteria, and three runs per prompt — and the March round will stay published alongside rather than being overwritten.

What this data lets you conclude, and what it does not

You can conclude, from the August run:

• Format compliance is solved across this cohort; stop weighting it.
• Claude Opus 5 builds the most interface per call and is the only model besides GPT-5.6 Luna Pro that never shipped an icon-free screen.
• GPT-5.6 Luna Pro has the best output-per-second ratio and by far the most predictable response time.
• Gemini 3.6 Flash is dense and fast but inconsistent on icons — four of ten screens had no vector icon at all.
• Llama 4 Maverick is not a candidate for finished screens, and it hard-failed once in ten calls.

You cannot conclude, from the August run:

• Which output looks better. We did not render and score them in this round.
• Whether high element counts mean better information architecture or worse over-nesting.
• Whether any model correctly built empty, loading, or error states. Element count bounds what can be present; it does not tell you what is.
• That the March panel scores apply to any August model. They do not — see the version table above.
• Anything about single-run variance. This run was one generation per prompt per model. The March protocol used three, and repeating it is what a variance claim would require.

That last one is a genuine limitation of the August data and we would rather name it than bury it. One run per cell tells you what a model typically does. It does not tell you how often it does something else.

Choosing by what you are actually doing

And the finding that outranks all of these: in the March study, the difference between a vague prompt and a specific prompt on the same model was larger than the difference between models on the same prompt. Before you switch models, fix your prompt. Our complete prompting framework for UI design covers the structure, and the prompt library has copy-paste starting points. If you want the short version, the UI prompt generator builds a structured prompt from a few dropdowns — free, no signup.

A note on graphic design

Two of the queries that lead people to pages like this one are "best LLM for visual design" and "best LLM for graphic design." We are only going to answer half of that.

This benchmark generated mobile app interfaces as HTML and CSS. It measured markup. It did not evaluate illustration, typography as a craft, poster layout, brand identity, or anything you would hand to a print shop. The models tested here cannot draw a logo — they emit markup, and an image model is a different tool with a different evaluation.

Ranking LLMs for graphic design on the strength of a UI markup test would be exactly the kind of unearned extrapolation this page exists to avoid. If someone shows you that ranking, ask what they rendered.

Where the model stops being the variable

Every number on this page is for raw model output: an LLM, one prompt, minimal system instructions, one call. That is a deliberately narrow test, and it is not how anyone ships a product.

The things that actually break in real design work are mostly not model capabilities:

• Cross-screen consistency. Each generation is independent. Ten calls give you ten unrelated colour scales unless something outside the model enforces one.
• Iteration. A benchmark scores the first output. Real work is the fourth.
• Design system adherence. No system prompt short enough to fit in a benchmark can carry your tokens, spacing scale, and component rules.

GenDesigns handles those three outside the model: a project-scoped theme generated once and injected into every screen, a chat loop for refinement, and routing so different tasks can use different models rather than forcing one to be good at everything. You can see the generation side at AI UI generator. GenDesigns is paid-only — plans start at $10/month, credits never expire, roll over month to month, and stay in your account if you cancel. There is no free tier. The UI prompt generator is genuinely free and works with any model on this page, including ones we do not sell.

Disclosure: GenDesigns is our product. The benchmark tests raw model APIs with no GenDesigns-specific tooling, and the model findings hold regardless of which tool you use. We have flagged our own limitations — including that we have not re-scored visual quality since March — as plainly as everyone else's.

If you are working with raw model output, the failure patterns are predictable enough to catalogue: common AI-generated UI mistakes and how to fix them. For the tooling layer rather than the model layer, see the best AI app design tools.

Frequently Asked Questions

What is the best LLM for UI/UX design?

It depends which half you mean. For the UI half — how much interface actually arrives and whether it includes real icons — Claude Opus 5 led our August 2026 run with a median 201 HTML elements and 24 inline SVG icons per screen, and never returned an icon-free screen in ten prompts. For the UX half — whether the information architecture is right — we have not scored the August cohort. Our March 2026 panel found layout logic and visual polish to be separate axes that rarely peak in the same model, but those scores belong to March models and we do not carry them forward.

Which LLM generates the most complete UI in one prompt?

Claude Opus 5, by a wide margin. It produced a median 201 HTML elements per screen against 175 for Gemini 3.6 Flash, 121 for Grok 4.5, 119 for DeepSeek V4 Pro, 109 for GPT-5.6 Luna Pro, and 35 for Llama 4 Maverick. On the most component-dense prompt, a food delivery listing with search, filters, and restaurant cards, Claude produced 295 elements while Llama produced 45.

Which LLM is fastest for UI generation?

Llama 4 Maverick had the fastest median latency at 15.8 seconds, but it also had the slowest rate of building interface at 1.41 elements per second — a fast response containing very little. On elements per second, Gemini 3.6 Flash led at 4.17 and GPT-5.6 Luna Pro followed at 4.07. GPT-5.6 Luna Pro was also the most predictable: all ten of its generations landed between 18.7 and 27.9 seconds, a 1.5x spread, versus 9.3x for DeepSeek V4 Pro and 21.2x for Llama 4 Maverick.

Which LLM produces the best icons in generated UI?

Claude Opus 5 emitted 223 inline SVG elements across ten prompts, against 119 for Grok 4.5, 89 for Gemini 3.6 Flash, 87 for GPT-5.6 Luna Pro, 76 for DeepSeek V4 Pro, and 13 for Llama 4 Maverick. Consistency matters more than the total: Claude Opus 5 and GPT-5.6 Luna Pro never returned a screen with zero inline SVG, while Gemini 3.6 Flash did so on four of ten prompts and Llama 4 Maverick on five of nine. We counted vector icons; we did not judge how good they looked.

Do LLMs still fail to follow UI generation instructions?

Not in this cohort. All six models were told to return one self-contained HTML document using Tailwind via CDN, target a 390px viewport, and output no markdown fences. Across 59 successful generations, every model complied on every instruction, every time. The one reliability problem was different in kind: Llama 4 Maverick failed one of ten generations outright with an Invalid JSON response from the provider.

What is the best LLM for graphic design?

This benchmark cannot answer that and neither can any UI generation test. We measured HTML and CSS output for mobile app screens. Graphic design means illustration, logos, print layout, and brand identity — output these models do not produce at all, since they emit markup rather than artwork. That work needs an image generation model and a completely different evaluation.

Is one run per prompt enough to rank models?

For the properties we measured, mostly yes, and we are explicit about the exception. Format compliance, element count, icon count, and latency are direct observations of what the model returned, and 59 generations across ten varied prompts is enough to see a 6x spread clearly. What one run per cell cannot tell you is run-to-run variance — how often a model does something atypical. The March protocol used three runs per prompt for exactly that reason, and any variance claim would need it repeated.

How do these results compare to the March 2026 benchmark?

They measure different things about different models. March scored visual quality, layout logic, code quality, component accuracy, and consistency using a three-person panel across 150 generations of Claude 4 Opus, GPT-4o, Gemini 2.5 Pro, DeepSeek V3, and Llama 4. August measured objective properties across 60 generations of an entirely new cohort. No model appears in both rounds, and Gemini moved tiers as well as versions, from 2.5 Pro to 3.6 Flash. The two sets of numbers are complementary, not comparable.

Can I reproduce this benchmark?

Yes. The harness is scripts/llm-ui-benchmark.ts in our repo, all six models were called through OpenRouter with the model IDs listed in the comparison table, and the ten prompts and the exact system prompt are published above. Latency will vary with provider load, and outputs are non-deterministic, so expect different absolute numbers and similar relative ordering.

Related Reading

• LLM UI Generation Benchmark 2026 — the full methodology, the March panel scores, and the August run in context
• AI Prompts for UI Design: The Complete Framework
• The Ultimate AI Design Prompt Library: 100+ Prompts
• Common AI-Generated UI Mistakes and How to Fix Them
• Best AI App Design Tools (2026)
• Free UI Prompt Generator