Benchmark·November 25, 2026·4 min read

Anthropic’s Claude models excel on real design briefs

Claude Fable 5.2, Opus 5.5 and Sonnet 5.5 ranked in the top 4 of 15 AI models on 33 real-world design tasks.

  1. 01Claude Fable 5.2 ranked 1st of 15 AI models at 1,713 Elo. Opus 5.5 ranked 2nd and Sonnet 5.5 4th.
  2. 02A Claude model leads each of the 7 rubric checks.
  3. 03Claude pages hold first place on 7 of the 12 briefs in the benchmark gallery.
Contra Labs
Contra Labs
Research

In the Frontier Design Benchmark, Claude Fable 5.2 placed 1st with an Elo of 1,713, Claude Opus 5.5 placed 2nd at 1,691, and Claude Sonnet 5.5 placed 4th at 1,667. Fable won 72.9% of its 683 blind comparisons, Opus 72.9% of 527, and Sonnet 69.1% of 440. Fable’s rating sits 26 points below the human-designed reference pages, at 1,739.

Leaderboard chart of Elo with 95% intervals for the human-designed reference pages and 15 AI models, with the three Claude models filled.
Elo with 95% intervals for the human-designed reference pages and the 15 AI models. The three Claude models are filled.
Share

Fable and Opus also earned the two highest Design Critique scores of any AI model, at 83.8 and 83.5. Sonnet scored 79.3. Frontier Design Benchmark leaderboard

A Claude model leads each of the 7 rubric checks below. On 5 of them, the three Claude models hold the top three places. The gap is widest on copy. Opus passed copy quality outright on 51.5% of its 33 outputs, while the median model outside the Claude family passed on 18.4%. On message clarity, Opus passed on 33.3% against a median of 4.9%.

The models also got the fundamentals right across the benchmark. Of their 93 outputs, 95% passed palette choice outright and 95% passed the check for touch target size and spacing.

Bar chart of the share of outputs that passed each rubric check outright for 15 AI models, with Claude models filled.
Share of outputs that passed each rubric check outright, for the 15 AI models (26 to 33 outputs per model). Claude models are filled, and the leading Claude model is named.
Share

Fable creates themes consistent with branding

Fable led the field in palette choice and icon clarity, and its palette passed outright on 32 of 33 rubrics. In addition, Bad Move’s brief asked for a crowdfunding page for an independent tabletop-game publisher. Backers in India needed to contribute once or monthly, pay with the methods they already use, and stay anonymous on a public Backer’s Wall. Fable included all three pieces of information on the first screen accurately.

Designers chose Fable in 20 of the 24 blind comparisons involving its Bad Move page, and it placed first on the brief. One designer explained the rationale of their preference:

It uses color combinations from the provided images with good color contrast. Easy to navigate, well organized layout. The typography looks similar to the one on the logo and creates a good balance.
Bad Move crowdfunding page built by Claude Fable 5.2.
Bad Move, built by Claude Fable 5.2.
Share

Sonnet makes clear content hierarchy

Sonnet led the rubric for type hierarchy and section order. It passed the hierarchy check outright on 25 of 27 outputs.

Broomwagon’s brief asked for a web app that lets small bicycle-repair shops publish a service menu without technical skills, starting with a landing page that drives sign-up.

Sonnet’s first screen reads top to bottom in one pass. A small label says who it is for: “A service menu for bike workshops.” The headline “Fix bikes. Not websites.” follows in heavy condensed capitals, with the second line set on a yellow band. One sentence then explains the offer: repair prices, opening hours and contact details become a menu customers can read on their phone, live in under ten minutes. To the right, a phone shows that menu for a sample shop, with a £12 puncture repair and a Book a service button. Designers chose Sonnet in 16 of the 18 blind comparisons involving its Broomwagon page. “B has the strongest brand personality,” one wrote, pointing to “a bold blue-and-lime ‘FIX BIKES. NOT WEBSITES.’ hero with a phone mockup of a published menu.” Another said Sonnet’s version “really took the given logo as the starting point for the design.” Sonnet placed first on the brief, and Opus placed second with 21 wins in 24 comparisons.

Broomwagon landing page built by Claude Sonnet 5.5, with a two-line headline and a phone showing a bike shop service menu.
Broomwagon, built by Claude Sonnet 5.5. A label, a two-line headline and one sentence lead to the sign-up button, with the finished menu shown on a phone.
Share

Claude at the Top of the Gallery

The benchmark gallery shows the top-rated AI page for 12 briefs. A Claude model holds first place on 7:

  • Claude Fable 5.2: Bad Move, NovaraAI, Ruckus
  • Claude Opus 5.5: Threatlane
  • Claude Sonnet 5.5: Broomwagon, Off You Go, Vellum

Each of the three models leads at least one brief, and on Threatlane, Broomwagon and Off You Go a second Claude model took the runner-up spot.

Gallery of the 7 benchmark briefs where a Claude page placed first.
The 7 gallery briefs where a Claude page placed first.
Share

Conclusion

The three Claude models won in different ways. Fable turned Bad Move’s requirements into a first screen a backer could act on, in colors that matched the game. Opus told Threatlane’s buyers what the product does in one sentence, then answered their next three questions. Sonnet gave Broomwagon a first screen you can read from top to bottom without hunting for the point.

They also share a signature. Claude models build the page out of the client’s own brand, turning a logo into a window, a lane or a pattern, and they draw their own flat illustrations when a brief needs one. Designers rewarded the result: Fable holds the top Elo among public models, all three Claude models sit in the top 4, and Claude pages hold first place on 7 of the 12 briefs in the gallery. Congratulations again to the Anthropic team.

How we ran this study → Methodology
Continue reading3 studies
All research
  1. November 18, 2026Benchmark
    Frontier Design: Evaluating AI on Real World Design TasksHow professional designers evaluated AI across 33 real-world design tasks.Read
  2. August 26, 2026Research
    How model performance shifted across four website design evalsHow model preferences shifted across ideation, mockup, and refinement.Read
  3. August 14, 2026Benchmark
    Claude models lead landing page design, but no model wins every design stageSix models built landing pages across three stages of creative work.Read

The world's leading independent human data & creative evaluation lab.

Powered by Contra.

The world's leading independent human data & creative evaluation lab.

Powered by Contra.