Who Designs Best? The Vote Winner vs the Designer’s Pick

Give the same design brief to several AIs at once, hide who made what, and let people pick the better one. Do that millions of times. Is the AI that wins those votes the one sitting on a designer’s desk?

Short answer: no. On September 24, 2026, OpenAI models led four of eight blind-vote categories, Meta models led two, and Anthropic and xAI models one each. Yet in a Q1 2026 designer survey, the most-used general AI was Claude at 78%, while ChatGPT, the leader a year earlier, slid to 65%.

The leaderboard kept moving. By October 6, Claude Opus 5.5 had taken #1 in 3D and data visualization, bringing Anthropic’s category wins to three. This piece reads the September 24 snapshot and flags what has changed since.

1. How does taste become a number?

Design Arena calls itself the largest crowdsourced benchmark for AI-generated design. Think of it as a blind taste test: the same prompt goes to several models, the results appear side by side with names hidden, and users vote for the better one. Votes from millions of users in 190+ countries are turned into an Elo rating, the same system used in chess.

The key idea: a rating gap is a win probability. A 100-point gap means the higher model wins about 64% of the time; 50 points, 57%; 25 points, 54%. Try it yourself:

Play with it and the picture sharpens. Pit September 24’s overall #1, Kimi K3 (1378), against #7, Claude Opus 5 (1331), and the leader’s expected win rate is only about 57%. In ten matchups, #7 would win about four. “Number one” sounds bigger than it is.

2. Category leaders: every task has a different winner

Beyond the overall board, Design Arena ranks models by category: websites, UI components, data visualization, 3D, mobile apps, image-to-website, slides, and image generation. Here were the category leaders on September 24:

OpenAI took four categories (UI component, data viz, 3D, image). Meta’s Muse Spark 1.3 led websites and image-to-website. Claude Fable 5.1 led slides, and Grok 4.6 led mobile apps, though only 5 points ahead of #2 Claude Fable 5 (1302 vs 1297). Scores are comparable only within a category, since each one has different models and votes. Tap the tabs to see each category’s top eight:

Three things stand out:

• The UI-component pack is tight. #1 GPT-6 Astra (1376) and #8 GPT-5.6 Sol (1350) are 26 points apart, so the leader’s expected win rate is just 53.7%. That is an expectation from the ratings, not a statistical test.
• 3D and slides had clear leaders. GPT-6 Astra led 3D by 50 points; Claude Fable 5.1 led slides by 34.
• Chinese models are all over the top ranks. Moonshot’s Kimi K3, the overall #1 on September 24, plus Xiaomi MiMo, Zhipu GLM and Alibaba Qwen made several top-eight lists.

3. Images: an OpenAI sweep

Text-to-image is the clearest race. Both the independent Artificial Analysis image arena and Design Arena’s image board put OpenAI’s GPT Image 2.5 on top, and both still did on October 6.

The interesting part is what’s missing. Midjourney is nowhere near the top of either arena. On Artificial Analysis, its best entry, Midjourney v7 Alpha, sat at #113 on October 6. Yet designer tool guides still call it the “aesthetic benchmark.” Blind votes judge one image; designers pay for taste that holds across a whole moodboard.

4. Survey: the AI designers actually use (Q1 2026)

Now look at the designer’s desk: “AI in Design 2026” from Designer Fund and Foundation Capital, which surveyed 900+ designers in 60+ countries. The survey ran in Q1 2026, about six months before the September 24 Design Arena snapshot.

The order flipped in a year. In this survey, the share using Claude rose from 52% to 78%, above ChatGPT (88% → 65%). Gemini was third at 48%; Perplexity fell from 34% to 13%.

Designers using AI at least weekly jumped from 54% to 91%, and the average number of AI tools per designer went from 3 to 7.

A second survey points the same way. In UX Tools’ “State of Prototyping Spring 2026” (1,478 designers and builders), Claude (50.8%) was the most-used weekly tool after Figma (82.6%).

ChatGPT (48.2%) ranked #3, and Claude Code (38.4%) #4, ahead of FigJam. Five of the top ten weekly tools are AI.

Source: UX Tools State of Prototyping Spring 2026 open data (1,478 respondents, multi-select). Bar length is relative to #1 Figma.

5. Why the vote winner isn’t the daily driver: three possible explanations

The two datasets answer different questions. A blind vote asks “which result do you like better?” A survey asks “what do you actually use?” These data alone cannot prove why the answers differ, but they point to three plausible explanations.

Product direction after the survey. After the Q1 survey closed, Anthropic launched Claude Design on April 17, 2026 as a research preview powered by Claude Opus 4.7, pitching design-system support and handoff to Claude Code. So Claude Design did not cause the Q1 usage numbers; it shows where products went afterwards.

Top two rows: AI in Design 2026. Bottom four: UX Tools State of Prototyping Spring 2026.

6. Voices from the field: praise and pain

Quotes from design leaders featured in the report share one theme: first drafts got dramatically faster, but adjusting spacing, fonts and brand rules exactly as intended is still human work.

In Figma’s “State of the Designer 2026” (906 designers), 91% say AI tools improve their designs and 89% say they work faster. Yet designers most often define craft as visual polish (58%). Speed went to AI; the bar for polish stayed with people.

7. So what should you use? Picks by task

Combining blind-vote rankings with real-world usage, here is a starting point, not a final answer. It is based on the September 24 rankings, with October 6 changes noted where a leader changed. Running the same brief through two or three models is still the surest test.

Caveats

Read this as a framework for judging, not a fixed ranking.

FAQ

Which AI model is best at design?

It depends on the task. On Design Arena as of October 6, 2026, Muse Spark 1.3 leads websites, GPT-6 Astra leads UI components and image-to-website, Claude Opus 5.5 leads 3D and data visualization, Claude Fable 5.1 leads slides, Grok 4.6 leads mobile apps, and GPT Image 2.5 leads images. Opus 5.5 also tops the overall (text-to-HTML) board, flagged as a preliminary result.

Which AI do working designers use most?

In AI in Design 2026 (surveyed in Q1 2026): Claude 78%, ChatGPT 65%, Gemini 48%. In the UX Tools survey, Claude (50.8%) was the #2 weekly tool after Figma.

How should I read Elo gaps?

A 100-point gap means about a 64% win rate for the higher model; 50 points, 57%; 25 points, 54%. Gaps of 20–30 points at the top are close to a coin flip. Compare Elo scores only within the same category, not across categories.

Why isn’t Midjourney in the rankings?

It is not near the top of the Artificial Analysis or Design Arena image boards. On October 6, the best Midjourney entry on Artificial Analysis, v7 Alpha, ranked #113. Designer tool guides still recommend it as an aesthetic reference for moodboards and style exploration.