We put four frontier models against the same ten landing page briefs, judged blind by 9 working designers. There was no single winner. On loose briefs GPT 5.6 Sol won 82% of its matchups. On briefs carrying a real design spec, Claude Fable 5 won and Sol finished last.
Three frontier models shipped inside a week, and two of them made the same promise. OpenAI said GPT 5.6 Sol was its strongest model yet, and that with only high-level direction it creates "tasteful, ergonomic, and functional interfaces." SpaceXAI said Grok 4.5 was its strongest model ever, and that it builds well-designed apps "even with minimal specification." Meta opened Muse Spark 1.1 to developers and priced it to undercut both.
Every launch post led with coding and agentic benchmarks. There is no equivalent benchmark for design.
So we asked designers. Ten landing page briefs, written in the format marketplace clients actually use. Four models: GPT 5.6 Sol, Claude Fable 5, Grok 4.5, and Muse Spark 1.1. Every model got an identical brief and an identical build instruction. Nine working designers judged the outputs blind, as live interactive pages. They picked a winner in each matchup, wrote up the strengths and weaknesses of every page on its own terms, and answered one more question: would you present this to a client?
Sol came out ahead overall, winning 63% of its matchups. Claude Fable 5 and Grok 4.5 finished close enough to call a tie, and Muse Spark 1.1 came last.
That ranking holds up right until you look at how each brief was written.
The spec decides the winner
Four of the ten briefs were loose. They gave a client, a goal, and a mood, and nothing else. Four were structured: section order and section contents, type family and weight, a grid, a palette, motion timings, and a short list of things not to do. Two sat in between.
Split the results along that line and the ranking inverts.
On loose briefs Sol scored 1703 on Elo ratings. On structured briefs it fell to 1455, behind Fable at 1569. Fable moved the opposite way, gaining 144 points the moment it was handed a spec. Muse Spark gained too. Grok barely moved.
All nine designers, independently, rated Sol higher on loose briefs than structured ones.
Sol placed first on every loose brief
Sol took first place on every loose brief in the set. Bramble 93%, Groundwork 81%, Rue Mistral 78%, Tidewater 78%. No other model placed first on a loose brief.
It beat each rival by roughly the same margin: Fable 81% of the time, Grok 83%, Muse Spark 83%. Sol was not being carried by one weak opponent.
On loose briefs, 61% of Sol's pages were judged client-ready. Across every other model combined, the figure was 31%. Nine professional designers, asked whether they would put their name on the work, said yes about twice as often for Sol as for the rest of the field.
Groundwork was the clearest case. An invite-only summit for climate-tech founders, briefed with a goal, a tone, and a single line of intent: make belonging feel earned. Sol won 81% of its matchups and seven of nine designers said they would present it, the highest client-ready rate of any loose page in the study.
What those designers praised was art direction. Across the loose pages they would present, colour and palette came up most, in more than three quarters of them, usually alongside the typography and the hero illustration. The word that recurs in the rationales is "captures": the page nails a tone the brief only gestured at. On Groundwork one designer called it "a typographic masterpiece that perfectly captures the urgent, credible, mission-driven tone."
The design has a strong personality and doesn't feel generic. The typography and color palette work well together to create a distinctive visual identity.
Sol's pages were 61% client-ready on loose briefs and 61% client-ready on structured ones. Fable's went from 31% to 72%, Muse Spark's from 28% to 56%, and Grok's from 36% to 47%.
The brief moves every model except Sol. It does the same quality of work whether you hand it a spec or a sentence. Sol wins on loose briefs because the rest of the field gets worse, not because Sol gets better.
A spec more than doubles what Fable can ship
Give Fable a specification and it more than doubles the share of its pages a designer would ship, from 31% to 72%.
On structured briefs it beat Sol 64% of the time, Grok 61%, and Muse Spark 64%. The advantage is uniform across opponents, so it belongs to Fable rather than to whatever Sol was doing.
Switchboard was the tightest brief we wrote: an eight-section running order, a named type stack, a grid, a four-colour palette, motion timings, and a tone spelled out as engineer to engineer. Fable read it, went dark, followed the running order, and wrote in the voice the brief asked for.
What designers praised in Fable's structured pages was fidelity rather than flair. More than two thirds of the ones they would present singled out the layout and structure, and better than a third credited the page specifically with following the brief, the highest rate of any model in any tier. The word that recurs here is "follows." On Switchboard one wrote that the page "flawlessly executes the requested layout structure and ordering." On The Undercurrent, another noted it "strictly follows all negative constraints, no images, no gradients, no SaaS cards."
All nine designers said they would present Fable's Switchboard page to a client. It is the only page in the study, out of forty, that every designer would ship.
This one doesn't break any rules.
I love it. Clean visual hierarchy, the text is spot on clean to read and feels engineer to engineer, love the darkmode.
On the same brief, Sol won one matchup out of twenty-seven. Constraints do not make Sol's pages worse, since its client-ready rate is flat. They make it unpredictable. Across the four loose briefs Sol's win rate barely moved, and across the structured ones its variance quadrupled. Fable's is the same on both.
They write better than they design
Copywriting was the most praised dimension for all four models. Designers quoted model copy back to us, unprompted, in the middle of critiques about spacing. Sol's Tidewater page: "The copywriting is phenomenal (e.g., 'Come in cold. Leave with a little more weather in you')." Grok handled the same brief's ban on biometrics with a section headed "We do not sell numbers."
Visual execution was the most criticized dimension for all four. One designer's verdict on Grok's Tidewater page holds the whole finding: "The copywriting on this output is exceptional. The visual execution completely fails the 'serene and elemental' tone."
Each model fails its own way.
Sol's is scale. On Switchboard it set the hero headline at 96 pixels against Fable's 56, wrapping to three lines and pushing the product visual below the fold. Four of nine designers led with it, and all four refused to send the page to a client. The brief named a type family and a type weight, and never named a type scale, so Sol chose one. It did the same with the palette, which named an ink colour and a paper colour without saying which was the background. Where a spec leaves a gap, Sol fills it. On a loose brief that is the entire point of the model.
Fable's is nerve. Handed a spec it produces the best pages in the study. Handed a mood it plays safe, and designers notice. One designer, on a page they still rated client-ready: "Very boring (but on other hand it has done what it was required)."
Grok is competent and forgettable on both kinds of brief. "The design is completely lifeless. There is a lack of risk that makes this design forgettable."
Muse Spark follows the plan and cannot compose the page. "The image on the right can't hold up against the weight of the copy on the left."
Designers also converged on the same tells, independently and across all four models. Em dashes, which one called a dead giveaway for AI writing. Small all-caps letter-spaced eyebrow headings above every section. Imagery that reads as placeholder no matter which model produced it. No coding benchmark measures any of it.
The takeaway
There is no best model here, only a question of how much you have already decided. If the art direction exists, Fable will execute it, and a specification more than doubles the share of its work a designer would put in front of a client. Hand it a mood instead and it plays safe, and safe reads generic. Sol is the opposite. It arrives with an opinion about how a landing page should look, and it holds that opinion at the same quality whether you give it a spec or a sentence, which is why writing Sol a careful brief does not get you a careful page.
Our tightest brief left two decisions open. It named an ink colour and a paper colour without saying which was the background, and it named a type family and a weight without naming a type scale. Sol filled both gaps itself, and was marked down for it. Fable filled them the way the nine designers expected. Neither model did anything wrong, which is the uncomfortable part. A spec with a hole in it is a spec the model will finish for you, and you will not learn which way it went until a designer tells you.
As raw capability converges, what separates a good output from a bad one keeps moving away from the model and towards the brief. That much we already knew from running Fable against Opus 4.8. The new part is that the models now have preferences about how that brief is written, and briefing one of them well means knowing which one you are talking to.
Methodology and limitations
Nine working designers in UI/UX and web design, sourced from Contra's network, evaluated ten landing page briefs built in the format of real marketplace client briefs. The clients were invented, so the set is unpublished and could not have appeared in any model's training data. Four briefs were scored loose, four structured, and two in between, across five dimensions of explicit design direction: type, spacing and grid, layout skeleton, palette, and references or constraints.
Each brief was run once through each model with the same standardized build instruction, producing self-contained HTML presented as a live, interactive artifact. Evaluators saw two outputs at a time, blinded and randomized, and picked a winner. They then assessed each output on its own for strengths, weaknesses, and client readiness. That produced 540 pairwise matchups and 360 written assessments. Elo Ratings are Bradley-Terry, converted to an Elo scale anchored at 1500, with 95% intervals from a cluster bootstrap over raters.
The two findings do not rest on equally firm ground. All nine evaluators rated Sol higher on loose briefs than structured ones (sign test, p = 0.004), and removing any single brief leaves the gap between 28 and 48 percentage points of win rate. Fable's structured advantage is larger in effect but noisier in support: seven of nine evaluators rated it higher on structured briefs than loose ones (p = 0.18), and it lost one of its four structured briefs outright.
We generated one page per model per brief, which is the ceiling on every claim above. The aggregate results average over ten briefs and nine raters and survive that, but individual cells carry confidence intervals roughly thirty points wide, so we treat them as evidence about the aggregate finding and never as a ranking. Switchboard is a single artifact, and we read it as an illustration of a pattern the full dataset supports rather than proof on its own.
The mid tier ran 39% loose, 36% mid, 59% structured, falling below loose. With only two mid briefs, both in difficult verticals, we read that as a vertical confound and do not lean on it. Brief structure and vertical are partially confounded throughout: our loose briefs skew taste-driven and our structured briefs skew execution-driven.
How we ran this study → Methodology