{"version":3,"file":"ResearchArticle.D1tPFXrt.mjs","names":["React.useRef","React.useState","_Fragment","React.useMemo","React.useCallback","x","React.Fragment"],"sources":["https:/framerusercontent.com/modules/kshvnQDlzqLIDcJri7zZ/E5K5uNrCS1RX6KvVolzT/ResearchArticle.js"],"sourcesContent":["// ResearchArticle.tsx — Framer code component for /research/[slug]\n//\n// Self-contained: all tokens, types, icons, and full article bodies are\n// inlined below. Drop this one file into Framer. No sibling imports.\nimport{jsx as _jsx,jsxs as _jsxs,Fragment as _Fragment}from\"react/jsx-runtime\";import*as React from\"react\";import{addPropertyControls,ControlType}from\"framer\";// ────────────────────────────────────────────────────────────\n// Tokens\nconst colors={cream:\"#F7F3EB\",off:\"#FBFAF6\",paper:\"#FFFFFF\",ink:\"#333330\",stone:\"#605A51\",fog:\"#9C968C\",sage:\"#A5BEBA\",bone:\"#DDD9C2\",mist:\"#C0CFCB\",line:\"#E9E5DA\",lineSoft:\"rgba(51,51,48,0.08)\",lineMed:\"rgba(51,51,48,0.15)\"};const fonts={sans:'\"GT Standard L Regular\", \"GT Standard L Regular Placeholder\", sans-serif',sansMed:'\"GT Standard L Medium\", \"GT Standard L Medium Placeholder\", sans-serif',serif:'\"Source Serif 4 Variable\", \"Source Serif 4 Variable Placeholder\", serif',mono:\"ui-monospace, SFMono-Regular, Menlo, monospace\"};const widths={prose:680,article:1e3,figure:1200,hub:1440};// ────────────────────────────────────────────────────────────\n// Container-width hook — measures the component's own root element via\n// ResizeObserver. Using window.innerWidth is wrong in Framer because the\n// component can render inside a smaller frame on a larger canvas (e.g.\n// the Mobile frame at 375px on a 1400px desktop preview).\nfunction useContainerWidth(){const ref=React.useRef(null);const[width,setWidth]=React.useState(1200);React.useLayoutEffect(()=>{const el=ref.current;if(!el)return;const rect=el.getBoundingClientRect();if(rect.width>0)setWidth(rect.width);if(typeof ResizeObserver===\"undefined\")return;const ro=new ResizeObserver(entries=>{for(const entry of entries){const w=entry.contentRect.width;if(w>0)setWidth(w);}});ro.observe(el);return()=>ro.disconnect();},[]);return[ref,width];}const fmtDate=iso=>new Date(iso+\"T00:00:00\").toLocaleDateString(\"en-US\",{year:\"numeric\",month:\"long\",day:\"numeric\"});function BrandAvatarFallback(){return /*#__PURE__*/_jsx(\"div\",{style:{width:44,height:44,borderRadius:\"50%\",background:colors.sage,border:\"1px solid rgba(51,51,48,0.12)\",display:\"flex\",alignItems:\"center\",justifyContent:\"center\",flexShrink:0},children:/*#__PURE__*/_jsx(\"svg\",{width:24,height:24,viewBox:\"0 0 24 24\",fill:colors.ink,xmlns:\"http://www.w3.org/2000/svg\",children:/*#__PURE__*/_jsx(\"path\",{d:\"M12 2 L13.5 10.5 L22 12 L13.5 13.5 L12 22 L10.5 13.5 L2 12 L10.5 10.5 Z\"})})});}function Icon({name,size=16,style}){const paths={arrow:/*#__PURE__*/_jsxs(_Fragment,{children:[/*#__PURE__*/_jsx(\"path\",{d:\"M5 12h14\"}),/*#__PURE__*/_jsx(\"path\",{d:\"m12 5 7 7-7 7\"})]}),arrowUp:/*#__PURE__*/_jsxs(_Fragment,{children:[/*#__PURE__*/_jsx(\"path\",{d:\"M7 17 17 7\"}),/*#__PURE__*/_jsx(\"path\",{d:\"M7 7h10v10\"})]})};return /*#__PURE__*/_jsx(\"svg\",{width:size,height:size,viewBox:\"0 0 24 24\",fill:\"none\",stroke:\"currentColor\",strokeWidth:1.5,strokeLinecap:\"round\",strokeLinejoin:\"round\",style:{display:\"inline-block\",verticalAlign:\"-2px\",flexShrink:0,...style},children:paths[name]});}// ────────────────────────────────────────────────────────────\n// Article data (full bodies)\nconst videoH2H={slug:\"video-head-to-head-jul26\",title:\"Frontier AI video models are nearly tied. None of them nail physics yet.\",type:\"Benchmark\",publishedDate:\"2026-06-29\",readMinutes:8,subtitle:\"A blind head-to-head of Seedance 2.0, Grok Imagine, Veo 3.1, and Adobe Firefly Video, judged by 12 professional video editors across 10 prompts.\",abstract:\"Seedance 2.0 finished first overall, but only by +0.07 over Grok Imagine on a 5-point scale. Across 10 prompts it won 4; Grok Imagine and Veo 3.1 won 3 each; Adobe Firefly Video won 0. The clearer pattern is specialization by shot type, and every model still breaks on physics.\",hero:false,coverImage:null,seoTitle:\"AI video head-to-head: Seedance 2.0 vs Grok, Veo, Firefly \\xb7 Contra Labs\",seoDescription:\"12 professional video editors judged Seedance 2.0, Grok Imagine, Veo 3.1, and Adobe Firefly Video across 10 prompts. The top three finished in a near-tie; all still fail on physics.\",ogImage:null,metrics:null,notionId:\"38bdf2b6-cc6b-80ac-af4e-dd1c9f79acc3\",updatedAt:\"2026-06-29T00:00:00Z\",body:[{type:\"keyFindings\",items:[\"Seedance 2.0 led almost every scalar dimension, but only narrowly; Grok Imagine and Veo 3.1 nearly matched it.\",\"The top three each specialized: Seedance on cinematic motion, Grok Imagine on people, Veo 3.1 on product and mood shots.\",\"Every model still showed physical glitches and realism failures, even on the highest-rated clips.\",\"Reference images lifted visual aesthetics by +0.43 but cut prompt adherence by -0.40.\",\"Adobe Firefly Video trailed the field and won none of the 10 prompts.\"]},{type:\"factbox\",label:\"How we ran it\",rows:[{label:\"Field\",value:\"AI video generation\"},{label:\"Sample\",value:\"10 prompts \\xb7 40 videos \\xb7 12 professional video editors\"},{label:\"Models\",value:\"Seedance 2.0, Grok Imagine, Veo 3.1, Adobe Firefly Video\"},{label:\"Method\",value:\"Blinded pairwise tournament + 1-5 scalar scoring across 5 dimensions\"}]},{type:\"p\",text:\"By mid-2026, AI video had turned from a demo category into a production-tool race, with four frontier models launching in close succession: Seedance 2.0, Veo 3.1, Grok Imagine, and Adobe Firefly Video. Seedance 2.0 arrived with the most drama: within days of its release in China, Axios reported a Disney cease-and-desist letter. Veo 3.1 led with native, synchronized audio generation, paired with better camera and narrative control. Grok Imagine claimed many #1 leaderboard spots, while Adobe Firefly Video trained exclusively on licensed content and focused on commercial use.\",parts:[\"By mid-2026, AI video had turned from a demo category into a production-tool race, with four frontier models launching in close succession: \",{text:\"Seedance 2.0\",href:\"https://seed.bytedance.com/en/seedance2_0\"},\", \",{text:\"Veo 3.1\",href:\"https://developers.googleblog.com/introducing-veo-3-1-and-new-creative-capabilities-in-the-gemini-api/\"},\", \",{text:\"Grok Imagine\",href:\"https://x.ai/news/grok-imagine-api\"},\", and \",{text:\"Adobe Firefly Video\",href:\"https://blog.adobe.com/en/publish/2025/02/12/meet-firefly-video-model-ai-powered-creation-with-unparalleled-creative-control\"},\". Seedance 2.0 arrived with the most drama: within days of its \",{text:\"release in China\",href:\"https://arxiv.org/abs/2604.14148\"},\", \",{text:\"Axios reported a Disney cease-and-desist letter\",href:\"https://www.axios.com/2026/02/13/disney-bytedance-seedance\"},\". Veo 3.1 led with native, synchronized audio generation, paired with better camera and narrative control. Grok Imagine claimed many #1 leaderboard spots, while Adobe Firefly Video trained exclusively on licensed content and focused on commercial use.\"]},{type:\"p\",text:\"To test which model generates the best professional-grade outputs, we recruited 12 professional video editors to compare 40 videos from the four models, also scoring them on prompt adherence, visual aesthetics, motion, and camera control. Our results showed that:\"},{type:\"list\",ordered:true,items:[\"Seedance 2.0 led almost every scalar dimension, but only narrowly. Grok Imagine and Veo 3.1 closely matched Seedance 2.0's performance, while Adobe Firefly Video trailed by a large margin.\",\"The top three models each performed best on different prompt types. Seedance 2.0 scored highest on dynamic cinematic motion, Grok Imagine on human subjects, and Veo 3.1 on low-motion product and mood shots.\",\"All models still showed physical glitches and realism issues. Even the strongest outputs exhibited various motion, continuity, and realism failures.\",\"Image references improved visual aesthetics, but reduced controllability. To prioritize the visual look and feel of a video, consider adding reference images to your prompts.\"]},{type:\"p\",text:\"We evaluated four video models across a set of 10 prompts, designed to test a set of quality aspects for video generation, including prompt adherence, model quality, camera shot quality, and visual aesthetics. Evaluators judged each output across five dimensions: physical glitches and issues, prompt adherence, motion quality, camera shot quality, and visual aesthetics. Then, they compared the video outputs in a tournament-style, pairwise ranking.\"},{type:\"p\",text:\"Seedance 2.0 finished first overall, but the margin was narrow, and the best-performing model changed by prompt type. The expert reviews showed where current video generation models do well and where they still lag.\"},{type:\"h2\",text:\"Seedance 2.0, Grok Imagine, and Veo 3.1 finished in a near-tie\"},{type:\"p\",text:\"Seedance 2.0 had the highest average score in the study, but it was not a dominant winner. Across 10 prompts, Seedance 2.0 won 4. Grok Imagine won 3. Veo 3.1 won 3. Adobe Firefly Video won 0.\"},{type:\"image\",slot:\"image1\",caption:\"Average score across all dimensions, on a 5-point scale. Seedance 2.0 leads the field, but by a fraction of a point over Grok Imagine and Veo 3.1.\"},{type:\"p\",text:\"That makes Seedance 2.0 the aggregate leader, but not a model that clearly separated from the rest of the field. Its advantage over Grok Imagine was only +0.07 on a 5-point scale. If you look into model performance on individual prompts, the results are also close: Seedance 2.0 won 4 out of 10 prompts, while Grok Imagine and Veo 3.1 both won 3 out of 10.\"},{type:\"image\",slot:\"image2\",caption:\"Prompt-level winners across the 10 prompts. Seedance 2.0 won 4, Grok Imagine and Veo 3.1 won 3 each, Adobe Firefly Video won none.\"},{type:\"p\",text:\"The average score can make the race look cleaner than it was. Seedance 2.0, Grok Imagine, and Veo 3.1 finished within a fraction of a point of each other, so the top of the leaderboard reads as a near-tie rather than a decisive win.\"},{type:\"h2\",text:\"Adobe Firefly Video trailed the field across every dimension\"},{type:\"p\",text:\"The dimension-level scalar results were more favorable to Seedance 2.0. It led in almost all measured dimensions, with the only exception being a -0.02 point loss to Veo 3.1 in Motion Quality.\"},{type:\"image\",slot:\"image3\",caption:\"Dimension-level scalar scores. Seedance 2.0 leads every measured dimension except Motion Quality, where Veo 3.1 edges it by 0.02.\"},{type:\"p\",text:\"Meanwhile, Adobe Firefly Video trailed by a large margin in all categories. It did not win any of the 10 prompts and consistently lagged behind the other models. In this dataset, Adobe Firefly Video was not competitive with the top three frontier video models.\"},{type:\"image\",slot:\"image4\",caption:\"Ranking distribution by model. Seedance 2.0 took the most 1st-place finishes but landed 3rd half the time; Grok Imagine ranked 1st or 2nd in 80% of prompts.\"},{type:\"p\",text:\"The ranking distribution tells a similar story. Seedance 2.0 and Grok Imagine reached nearly identical average win rates through opposite behaviors. Seedance 2.0 was the high-ceiling, more polarizing pick: it earned the most 1st-place finishes (4 of 10 prompts) and never finished last, but landed in 3rd place half the time, with little in between. Grok Imagine was the more consistent top-half performer, ranking 1st or 2nd in 80% of prompts, though it occasionally fell to 4th (2 of 10 prompts).\"},{type:\"h2\",text:\"Image references improved visual aesthetics, but reduced controllability\"},{type:\"p\",text:\"Adding reference images to prompts improved visual aesthetics, but reduced controllability. Average visual aesthetics increased from 3.04 to 3.47, a gain of +0.43. But prompt adherence fell from 3.07 to 2.67, a drop of -0.40. The reference image helped the outputs look better, but it also made the models less responsive to the written prompt.\"},{type:\"table\",headers:[\"Dimension\",\"Text only\",\"Text + image\",\"Change\"],rows:[[\"Visual Aesthetics\",\"3.04\",\"3.47\",\"+0.43\"],[\"Prompt Adherence\",\"3.07\",\"2.67\",\"-0.40\"]]},{type:\"p\",text:\"That tradeoff showed up at the model level too. Three of four models got worse. Seedance 2.0, the text-to-video leader, fell the most, dropping -0.72. The only model that improved was Adobe Firefly Video, which gained +0.57. The reference image appears to improve surface-level aesthetics while making the model harder to steer.\"},{type:\"image\",slot:\"image5\",caption:\"Effect of adding a reference image, per model. Three of four models got worse at following the prompt; only Adobe Firefly Video improved, gaining +0.57.\"},{type:\"p\",text:\"For Adobe Firefly Video, adding images to the prompt noticeably improved the Elo rating of the model's output. This suggests image references can raise output quality for a model that is further from the frontier.\"},{type:\"callout\",text:\"For video creatives: when look and feel are your top priority, include reference images. But understand it can make the model harder to steer. Keep the written prompt simple and let the image carry the style.\"},{type:\"h2\",text:\"All models still suffered from physical glitches and realism issues\"},{type:\"p\",text:\"Even when the outputs were strong, all models still showed physical glitches and realism issues. When asked to identify these issues in the videos, most participants were able to identify multiple issues in all of the generated videos.\"},{type:\"p\",text:\"Even the strongest outputs were not clean. The two clips below were among the best-rated in the study, yet evaluators still flagged multiple issues in each (representative, not comprehensive):\"},{type:\"video\",slot:\"video1\",caption:\"Veo 3.1, 61% win rate. Prompt: a professional news anchor in a grey suit at a modern studio desk delivers a short news line to camera with natural lip movement, static medium shot, synchronized audio.\"},{type:\"list\",ordered:false,items:[\"Camera movement (0:03-0:08): 'slight upward camera movement, which would not happen in a news programme either.'\",\"Hands (0:04-0:06): 'The pen on his right hand glitches and distorts as he moves his hands.'\",\"Body and edges (0:03-0:05): 'Very slight edge issues with his shoulders, but nothing someone like me (a VFX sup) would notice.'\"]},{type:\"p\",text:\"The second clip, despite topping its prompt, drew the same kind of physics complaints (representative examples, not comprehensive):\"},{type:\"video\",slot:\"video2\",caption:\"Seedance 2.0, 78% win rate. Prompt: a direct overhead bird's-eye view of a busy four-way intersection, pedestrians crossing and cars moving through as the signal changes, camera slowly descending and rotating overhead.\"},{type:\"list\",ordered:false,items:[\"Duplicate vehicle (0:01.967): 'On the left-hand side, a duplicate black car right behind the original one.'\",\"Traffic (full video): 'All cars are going at the same time, causing wrecks;' 'No visible signal or lights changing for traffic.'\",\"Vehicle glitch (0:01-0:07): 'A black car entering from the left side of the intersection appears unusually elongated.'\",\"Pedestrian glitch (0:06-0:08): 'several instances where pedestrians on the sidewalks merge into one another.'\"]},{type:\"p\",text:\"This was the shared failure mode across the benchmark. Even though the models have improved on overall scene, mood, and visual style, they still break down in the details that make a video physically believable. Physical realism remains a major weakness even for the best models. If you are a creative professional generating videos with AI, make sure you plan for manual VFX cleanup on hands and props, or keep clips short to limit drift.\"},{type:\"h2\",text:\"Without a clear winner, the models showed specialization\"},{type:\"p\",text:\"Because the top three models were close, the clearer pattern is specialization by prompt type. Seedance 2.0 performed best on dynamic, high-energy cinematic shots. Veo 3.1 and Grok Imagine performed well on static or slower movement shots. Grok Imagine specialized in human details, while Veo 3.1 worked great for product shots and object showcase.\"},{type:\"p\",text:\"Specifically, Seedance 2.0 was strongest when the prompt required dynamic action or cinematic motion. It performed best on the alley chase, night market, and wizard prompts, and was the least-bad option on traffic.\"},{type:\"video\",slot:\"video3\",caption:\"Seedance 2.0, rated well above the field. Prompt: a person in a dark hoodie sprints down a narrow rain-soaked alley at night, handheld camera chasing at a tilted Dutch angle, neon reflections, teal-and-orange grade, fast motion.\"},{type:\"p\",text:\"Grok Imagine stood out on human performance, performing best on the barista prompt. Its strength in this dataset was in scenes centered on people's facial expressions and body movement.\"},{type:\"video\",slot:\"video4\",caption:\"Grok Imagine, rated highest among the models. Prompt: a young barista with curly red hair, freckles, and a green apron wipes a glass cup behind a cafe counter, warm tungsten lighting, static medium close-up, cinematic film-grain look.\"},{type:\"p\",text:\"Veo 3.1 performed best on lower-motion, mood-driven, product-oriented prompts. It won the mug, sneaker, and candle prompts, where lighting, object presentation, and atmosphere mattered more than high-energy motion.\"},{type:\"video\",slot:\"video5\",caption:\"Veo 3.1. Prompt: a single lit candle on a dark wooden table in an otherwise black room, warm glow on the wax, a faint wisp of smoke curling upward, camera holds still.\"},{type:\"p\",text:\"Noticeably, Veo 3.1 was the lone exception to Seedance 2.0's lead. It received slightly higher scalar rating than Seedance 2.0 on Motion Quality and tied it on Visual Aesthetics. That complicates the dimension-level story: the three frontier models performed closer than the aggregate leaderboard makes it look.\"},{type:\"p\",text:\"The main takeaway is that the best model depended on the type of video being generated. Seedance 2.0 had the strongest average, but the prompt-level winners showed that model choice should change by workflow pattern.\"},{type:\"h2\",text:\"Conclusion\"},{type:\"p\",text:\"Seedance 2.0 was the strongest overall model in this video head-to-head study, but it was not a runaway winner.\"},{type:\"p\",text:\"It won the highest average score and led in almost all measured dimensions. But across 10 prompts, it won only 4. Grok Imagine and Veo 3.1 each won 3. Seedance 2.0's margin over Grok Imagine was only +0.07, and its head-to-head pattern against Veo 3.1 was effectively split.\"},{type:\"p\",text:\"The more useful conclusion is that the models specialize. Seedance 2.0 was strongest for high-energy cinematic motion. Grok Imagine was strongest for human performance and identity-heavy scenes. Veo 3.1 was strongest for low-motion product, mood, and lighting shots. Adobe Firefly Video trailed the field in this dataset.\"},{type:\"p\",text:\"The benchmark also shows that AI video is still constrained by physical glitches, realism issues, and controllability tradeoffs. Image references improved aesthetics, but reduced prompt adherence.\"},{type:\"p\",text:\"The takeaway: Seedance 2.0, Grok Imagine, and Veo 3.1 are effectively neck and neck, and the best model depends on the shot.\"},{type:\"h2\",text:\"Methodology\"},{type:\"p\",text:\"Twelve professional video editors sourced from the Contra Labs network evaluated 40 AI-generated video outputs across 10 prompts. The prompt set was designed to cover a range of creative video use cases: dynamic cinematic motion, human performance, product and object shots, mood-driven scenes, stylized visual effects, and physically complex scenes where temporal consistency and realism are difficult to maintain.\"},{type:\"p\",text:\"The study compared four video generation models: Seedance 2.0, Grok Imagine, Veo 3.1, and Adobe Firefly Video. Each prompt was run through all four models, producing four outputs per prompt. Five prompts were text-to-video tasks, generated from a written prompt only. Five prompts were image-to-video tasks, generated from a reference image plus a written prompt. For the image-to-video condition, reference images were generated and iterated using ChatGPT Image until they met the quality bar needed for the evaluation. Outputs were presented blind as Output A-D, so evaluators judged the videos without seeing which model produced each clip.\"},{type:\"p\",text:\"The prompt set spanned several creative use cases. Three examples:\"},{type:\"table\",headers:[\"Sample Prompts\",\"What it tested\"],rows:[[\"A white ceramic mug tips over on a polished marble kitchen counter and spills hot black coffee, the liquid spreading and dripping over the edge as steam rises. Soft morning light from a window on the left. Static medium shot at counter height, shallow depth of field, warm natural color grade.\",\"Liquid physics, steam, object interaction, lighting, and static composition.\"],[\"A slow crane shot rising from street level up over a crowded Asian night market at dusk, revealing rows of glowing red lanterns and food stalls with steam, people walking below. Begins as a low eye-level shot and ends on a high wide establishing shot. Saturated neon color palette, rich contrast.\",\"Smooth crane movement, shot-scale transition, crowd consistency, lighting, and composition.\"],[\"A robed wizard raises one hand and conjures a glowing orb of magical energy that swells and crackles with arcs of blue-and-violet light, streams of glowing particles and embers swirling outward and trailing through the air as wisps of smoke curl around the hand. A dark stone chamber lit only by the spell's shifting glow. Stylized, painterly fantasy look with dramatic rim lighting, rich saturated color, slow-motion energy. Static medium shot, eye level.\",\"Particle effects, stylized lighting, color consistency, motion, and fantasy visual coherence.\"]]},{type:\"p\",text:\"Evaluators completed three tasks for each prompt. First, they identified visible glitches and physical realism issues, describing what was wrong and what should change. These annotations captured artifacts such as warping, flicker, hallucinated details, object deformation, unnatural motion, camera problems, prompt mismatches, and failures of physical plausibility. Second, evaluators rated each output individually on four scalar dimensions using a shared 1-5 rubric: Prompt / Reference Adherence, Motion Quality, Camera & Shot Quality, and Visual Aesthetics. Third, evaluators completed a blinded pairwise tournament across the four outputs for the same prompt, choosing the stronger output in each head-to-head comparison.\"},{type:\"p\",text:\"We aggregated the results across prompts to compare average model performance, prompt-level winners, dimension-level scores, ranking distribution, and derived pairwise win rates. Qualitative comments were used to interpret why outputs succeeded or failed, especially around motion stability, physical realism, visual polish, prompt adherence, and production-readiness.\"},{type:\"h2\",text:\"Limitations\"},{type:\"p\",text:\"This study is limited to 10 prompts, four models, and 12 professional video editors, so the results should be read as directional rather than definitive. Model rankings may change with different prompts, different model versions, different generation settings, longer or shorter clips, or a different evaluator panel.\"},{type:\"p\",text:\"The comparison reflects first-pass output quality under the study setup, not each model's ceiling under extensive prompt iteration, manual selection, editing, or post-production. The prompt set intentionally covered several creative use cases, but it does not represent the full range of video generation workflows, such as long-form narrative continuity, multi-shot editing, brand-safe production pipelines, audio-first generation, or complex client revision cycles.\"},{type:\"p\",text:\"The evaluation also emphasizes visible video quality rather than every production constraint. Editors judged prompt / reference adherence, motion, camera work, visual aesthetics, and physical glitches, but the study did not separately score factors such as audio quality, editability, latency, cost, licensing, safety controls, or downstream integration into professional tools.\"},{type:\"p\",text:\"Evaluator judgment is inherently subjective. The shared rubric helped anchor scoring, but professional taste, tolerance for artifacts, and expectations for production-ready video can vary. We therefore place more weight on patterns that repeated across prompts, models, and evaluator comments than on any single score or individual matchup.\"}]};const designCrit={slug:\"design-crit\",title:\"Introducing Design Crit: we taught AI to judge design like a designer.\",type:\"Research\",publishedDate:\"2026-06-17\",readMinutes:7,subtitle:\"Ten professional designers ranked four frontier image models across nine dimensions of real design work. The models can make the work. Nothing on the market could reliably judge it, until we trained on the right data.\",abstract:\"Design Crit is a designer-annotated preference dataset for AI-generated graphic design, recording one rating per design-quality criterion instead of one verdict per image. No off-the-shelf judge cleared 55% agreement with a five-designer panel, where a human designer hits 74.1%. A small model trained on Design Crit reaches 0.611, closing roughly 46% of the gap between chance and the human ceiling, and on the hardest 3-2 splits it already ties a human.\",hero:false,coverImage:null,seoTitle:\"Design Crit: teaching AI to judge design like a designer \\xb7 Contra Labs\",seoDescription:\"Ten designers ranked four frontier image models across nine criteria. No off-the-shelf judge beat 55% agreement with the panel. A model trained on Design Crit closes 46% of the gap to human.\",ogImage:null,metrics:null,body:[{type:\"keyFindings\",items:[\"Ten designers ranked four image models on nine criteria, one rating per axis.\",\"No off-the-shelf judge beat 55% agreement. The best hit 54.3%; a human hits 74.1%.\",\"Scaling doesn't help: Qwen3-VL at 4B, 8B, and 32B all stall near 51–54%.\",\"One in ten finished designs carried a major hallucination.\",\"Trained on Design Crit, a model reaches 61.1% agreement with the panel and ties a human on the hardest splits.\"]},{type:\"factbox\",label:\"How we ran it\",rows:[{label:\"Models\",value:\"FLUX.2 max, GPT Image 1.5, Nano Banana 2, Seedream 5.0 Lite (shown blind, code-named)\"},{label:\"Raters\",value:\"10 professional designers from Contra's network, split into aesthetics and fidelity cohorts of 5\"},{label:\"Method\",value:\"9 criteria \\xb7 80 prompts each \\xb7 1,600 ratings per criterion \\xb7 all 6 pairwise comparisons aggregated into 4-way rankings\"},{label:\"Judges tested\",value:\"9 pre-trained systems (HPSv2.1, PickScore-v1, LAION-Aesthetic-V2 + 6 open-weight VLMs)\"}]},{type:\"video\",slot:\"video1\",caption:\"Design Crit (Criteria-Resolved Image Taste), a Lica \\xd7 Contra collaboration.\"},{type:\"p\",text:\"Everyone keeps talking about taste. But you can't improve what you can't measure. So we measured it. Design Crit is a dataset of ten professional designers ranking four frontier image models across nine dimensions of real design work. The models can produce the work. Nothing on the market can reliably judge it. The good news is that what they're missing can be learned.\"},{type:\"p\",text:'Text-to-image models have matured from research demos into deployed design tools, shipping posters, social posts, UI mockups, and logos straight into production. But the preference data that trains and grades them was collected on photo-style, scene-based generation, where a single \"which one is better?\" verdict captures almost everything that matters, like sharpness and prompt alignment.'},{type:\"p\",text:\"Designers, however, don't judge a design on a single characteristic, and design doesn't collapse into one axis. A graphic design can nail the spatial structure and butcher the color intent. Another can satisfy the brief and break the typographic hierarchy. Both get the same overall thumbs-up, for completely different reasons. The signal designers actually use lives in the dimensions a single label averages away.\"},{type:\"p\",text:\"So we built the missing layer. Design Crit (Criteria-Resolved Image Taste) is a designer-annotated preference dataset for AI-generated graphic design. It records one rating per design-quality criterion instead of one verdict per image. That lets a system score a design on every axis at once, make sharper calls than a single label allows, and in time learn to weigh those axes the way a designer would.\"},{type:\"h2\",text:\"10 designers, 4 models, 9 ways to be wrong\"},{type:\"p\",text:\"We put four current text-to-image models head to head. FLUX.2 max, GPT Image 1.5, Nano Banana 2, and Seedream 5.0 Lite, all shown to raters behind blind code-names so no one was anchoring on a brand.\"},{type:\"p\",text:\"Ten professional designers, recruited through Contra's network of creative experts, split into two cohorts of five. One cohort judged aesthetics, covering overall preference, mood and tone, visual hierarchy, color harmony, and typographic craft. The other judged description fidelity, covering overall preference, color accuracy, spatial accuracy, and whether the text the brief asked for was actually rendered.\"},{type:\"image\",slot:\"image1\",caption:\"One of the AI-generated graphic designs in the set. Designers rated work like this one axis at a time, not with a single overall verdict.\"},{type:\"p\",text:\"We narrowed the nine criteria from a longer list through pilot studies and interviews, building on Contra's Human Creativity Benchmark and keeping the axes designers consistently treated as separate.\",parts:[\"We narrowed the nine criteria from a longer list through pilot studies and interviews, building on \",{text:\"Contra's Human Creativity Benchmark\",href:\"https://contralabs.com/human-creativity-benchmark\"},\" and keeping the axes designers consistently treated as separate.\"]},{type:\"p\",text:\"Nine criteria, 80 prompts each, every prompt scored by five designers across four models, for 1,600 ratings per criterion. Designers ran all six pairwise comparisons per prompt, and we aggregated those into strict four-way rankings. On the two overall-preference tracks they also flagged every image for hallucination.\"},{type:\"p\",text:\"These ratings come from working professionals, and they fill the part of the stack everyone training these models has been flying blind on.\"},{type:\"h2\",text:\"Design is subjective, but real and consistent\"},{type:\"p\",text:\"A fair question is whether designers just don't agree, leaving no signal to learn. We tested that directly, checking whether designer agreement exceeds what random raters would produce, against exact null distributions.\"},{type:\"p\",text:'They do agree. Designers agree on good design about as much as people agree on their favorite movie, and less than crowds agree on which photo is sharper. And the way they disagree is healthy. They share a rough sense of \"good,\" with some personal variation on top. There are no rival camps with opposite taste. That\\'s exactly the kind of pattern a model can learn.'},{type:\"image\",slot:\"image2\",caption:\"Where design sits on the subjective-to-objective scale: more agreement than favorite-movie picks, less than judging which photo is sharper.\"},{type:\"p\",text:\"But that agreement isn't even across the dimensions. We measured how often designers landed on the same call for each criterion, and the gap is wide. The axes you can check against the brief draw the tightest agreement, whether the requested text actually rendered, whether the layout is right, whether the colors match what was asked for. The axes that come down to pure feel draw the least, with color harmony at the bottom.\"},{type:\"p\",text:\"The clearest read is the matched pairs. Designers agree far more on whether the right text rendered than on whether the type is well set. They agree more on whether the requested color appeared than on whether the colors sit well together. Same subject each time, the checkable version high, the felt version low. The signal is real on every axis. It just gets noisier the more the call comes down to taste.\"},{type:\"image\",slot:\"image3\",caption:\"Designer agreement by criterion (Kendall's τ). Checkable axes like text rendering rank highest; felt axes like color harmony rank lowest.\"},{type:\"h2\",text:\"No off-the-shelf judge beats a coin flip\"},{type:\"p\",text:\"So the signal is real. The question is whether anything on the market can read it. We benchmarked nine pre-trained systems as design judges. Three were dedicated preference and aesthetic scorers (HPSv2.1, PickScore-v1, LAION-Aesthetic-V2) and six were open-weight vision-language models prompted to pick the better image.\"},{type:\"p\",text:\"Not one cleared 55% agreement with the five-designer majority. Chance is 50%. The best system, HPSv2.1, was trained on more than 640,000 human image comparisons, and it lands at 54.3%. LAION-Aesthetic-V2 actually scores below chance. A human designer agrees with the panel 74.1% of the time. Every machine judge sits in the dead zone just above a coin flip.\"},{type:\"image\",slot:\"image4\",caption:\"Agreement with the five-designer majority. Every off-the-shelf judge clusters just above chance; a human designer sits far above at 74.1%.\"},{type:\"h2\",text:\"You can't compute your way to taste\"},{type:\"p\",text:\"Scaling doesn't move the number. Qwen3-VL at 4B, 8B, and 32B all land between 51% and 54%. The reason is a trade-off. Bigger models carry less position bias, so their pick barely changes when you swap the order of the two images. They're more internally consistent. But that consistency buys no accuracy. On the calls they commit to, the bigger models are no better, and a little worse. Smaller models lean on position more, yet the calls they do commit to are sharper. The more a model leans on position, the better it judges when it doesn't (Spearman ρ = +0.94), so the two effects cancel and the total never moves. The bottleneck is the data.\"},{type:\"image\",slot:\"image5\",caption:\"Model size vs. agreement for Qwen3-VL at 4B, 8B, and 32B. Scaling the model leaves accuracy flat between 51% and 54%.\"},{type:\"h2\",text:\"One in ten designs hallucinates something the prompt never asked for\"},{type:\"p\",text:\"Reading design isn't the only place the models slip. While they ranked the designs, the same designers flagged every image on the overall-preference tracks for hallucination, meaning elements that drifted from the brief or had nothing to do with it. Across 1,600 flags per cohort, about 55% came back clean, 35% minor, and 10% major. One in ten finished designs carried a major hallucination, something the prompt never asked for. These are the kinds of failures a designer catches at a glance and the model that made them does not.\"},{type:\"image\",slot:\"image6\",caption:\"Hallucination flags across 1,600 designs per cohort: 55% clean, 35% minor, 10% major.\"},{type:\"h2\",text:\"Train on the data, and half the gap disappears\"},{type:\"p\",text:\"Then comes the turn. We trained a small pairwise-difference head on top of a frozen vision-language encoder, with no fine-tuning of the backbone, a deliberately modest model, directly on Design Crit.\"},{type:\"p\",text:\"It reaches 0.611 agreement with designers. That closes roughly 46% of the entire gap between a coin flip (0.500) and the human ceiling (0.741), and it's the first configuration in our sweep to clear the noise floor that standard regularization couldn't budge. The lesson from the benchmark holds in reverse. The signal was always there. It just had to be trained on the right data instead of borrowed from photo preference.\"},{type:\"image\",slot:\"image7\",caption:\"A small model trained on Design Crit reaches 0.611, closing about 46% of the gap between chance and the human ceiling.\"},{type:\"h2\",text:\"On the genuinely hard calls, it already ties a human\"},{type:\"p\",text:\"Roughly half of all pairwise comparisons are 3-2 splits, cases where the designers themselves are nearly evenly divided and even a perfect predictor is partly guessing. Those are the calls that actually test judgment.\"},{type:\"p\",text:\"On exactly those cases, a model trained on Design Crit scores 0.602 against a human ceiling of 0.600. When the panel splits three to two, even one designer only agrees with the majority three times in five, and the model now matches that. The gap to human agreement stays wide on the cases designers find easy. On the ones they find hardest, it shrinks to almost nothing.\"},{type:\"image\",slot:\"image8\",caption:\"On the hardest 3-2 splits, the Design Crit model (0.602) matches the human ceiling (0.600).\"},{type:\"h2\",text:\"Why it matters\"},{type:\"p\",text:\"Design Crit (Criteria-Resolved Image Taste) lets you build a decision layer for design generation. Its criterion-level structure means you can route between generators by what a job actually needs. Pick the model that's strongest on typography for a logo, or on spatial accuracy for a layout, rather than trusting one aggregate score. The same structure works as supervision for training preference judges and reward models that optimize for specific design dimensions rather than a blurry average.\"},{type:\"p\",text:\"The headline finding is blunt. AI can generate design, but it cannot yet reliably tell good design from bad, and no amount of scale fixes that on its own. The hopeful finding is that the missing signal, taste, is real and can be learned from expert data. That's the layer Contra's network is built to provide.\"},{type:\"p\",text:\"Our first Design Crit dataset “TASTE” is published on arXiv: arxiv.org/abs/2605.20731\",parts:[\"Our first Design Crit dataset “TASTE” is published on arXiv: \",{text:\"arxiv.org/abs/2605.20731\",href:\"https://arxiv.org/abs/2605.20731\"}]},{type:\"h2\",text:\"Limitations\"},{type:\"p\",text:\"The sample is small. Each prompt was rated by five designers, enough to measure agreement and rule out noise but not enough to be sure of any single comparison. Each criterion used its own set of 80 prompts, so no design was ever rated on two criteria at once. That kept each rating clean, but it means we can't see how one designer weighs color against typography on the same design, because no one judged the same design on both. Every prompt was in English, so taste across languages isn't captured here. And the nine criteria cover a lot, but not everything. Accessibility, brand consistency, motion, and audience fit are all natural axes to add as the work grows.\"},{type:\"h2\",text:\"Future research\"},{type:\"p\",text:\"The obvious next move is to run this at a wider scale, with more designers per prompt, more languages, and the criteria expanded. Rating the same designs on every axis might reveal how designers balance color against hierarchy or fidelity against feel, the trade-offs a single score hides.\"},{type:\"p\",text:\"So far we've shown the signal can be learned as a judge. The open question is whether it makes models better designers. Training a generator against these per-dimension rewards could push on typography or color directly, then show whether the work actually improves.\"}],notionId:\"375df2b6-cc6b-802e-af18-cc89e33ebf18\",updatedAt:\"2026-06-17T00:00:00Z\"};const codeEditorBattle={slug:\"code-editor-design-battle\",title:\"Cursor took 60% of head-to-heads. Claude Code took 63% of client meetings.\",type:\"Benchmark\",publishedDate:\"2026-05-29\",readMinutes:6,subtitle:\"Four coding tools, 24 outputs, five working designers. The tool designers preferred to look at and the tool they'd put their name on turned out to be different.\",abstract:\"Cursor won the head-to-head tournament at 60%. Claude Code won the client-readiness vote at 63%. Antigravity and Codex split the rest. Side-by-side preference and client readiness run on different logic, and no tool cleared the bar on mobile.\",hero:false,coverImage:null,seoTitle:\"Cursor vs. Claude Code vs. Codex vs. Antigravity, judged by working designers \\xb7 Contra Labs\",seoDescription:\"We ran 24 outputs from Cursor, Claude Code, Codex, and Antigravity past five working designers. Cursor took 60% of head-to-heads. Claude Code took 63% of client meetings.\",ogImage:null,metrics:null,body:[{type:\"keyFindings\",items:[\"Cursor won the head-to-heads: 60% pairwise win rate, top Elo (1575).\",\"Client readiness flipped it: Claude Code 63%, Antigravity 53%, Cursor 47%, Codex 30%.\",\"Pairwise wins reward surface polish; client readiness rewards defensible work.\",\"No tool led every format: Cursor took landing pages, Antigravity web apps.\",\"Mobile was weakest: no tool cleared a 3.0 composite out of 5.\"]},{type:\"factbox\",label:\"How we ran it\",rows:[{label:\"Tools\",value:\"Cursor (composer-2.5), Claude Code (claude-sonnet-4-6), OpenAI Codex (gpt-5.5-medium), Google Antigravity (gemini-3.5-flash-high)\"},{label:\"Briefs\",value:\"6 fictional client briefs \\xb7 2 landing pages, 2 mobile apps, 2 web apps\"},{label:\"Outputs\",value:\"24 live HTML pages \\xb7 evaluated as interactive artifacts\"},{label:\"Evaluators\",value:\"5 working designers from Contra's top-earning talent\"}]},{type:\"image\",slot:\"image1\",caption:\"Four coding tools, 24 outputs, five working designers. Where their pairwise picks and their client-meeting picks parted ways.\"},{type:\"p\",text:\"The SF Standard ran a headline: “'Engineer' is so 2025. In AI land, everyone's a 'builder' now.” Figma's own 2025 report found 56% of non-designers are now performing design-centric tasks. The conclusion seems obvious. The tools have caught up, the role is dissolving, anyone can ship beautiful interfaces.\",parts:[\"The SF Standard ran a headline: \",{text:\"“'Engineer' is so 2025. In AI land, everyone's a 'builder' now.”\",href:\"https://sfstandard.com/2026/03/05/engineer-2025-ai-land-everyone-s-builder-now/\"},\" Figma's own 2025 report found \",{text:\"56% of non-designers are now performing design-centric tasks\",href:\"https://www.figma.com/blog/2025-shifting-roles-report/\"},\". The conclusion seems obvious. The tools have caught up, the role is dissolving, anyone can ship beautiful interfaces.\"]},{type:\"p\",text:\"The interfaces all look the same. Give the same prompt to any of these tools and what comes back is the same left-rail sidebar, the same rounded card grid, the same indigo accent that's become the unofficial flag of AI-generated interfaces. Designers have started calling it “product slop.” The tools give you the average of what they've seen because nothing in the input asks them not to.\",parts:[\"The interfaces all look the same. Give the same prompt to any of these tools and what comes back is the same left-rail sidebar, \",{text:\"the same rounded card grid, the same indigo accent\",href:\"https://www.banani.co/blog/vibe-designing\"},\" that's become the unofficial flag of AI-generated interfaces. Designers have started calling it \",{text:\"“product slop.”\",href:\"https://medium.com/@Rythmuxdesigner/the-anti-vibe-coding-movement-why-designers-are-rejecting-ai-generated-mediocrity-in-2026-b241ce60a369\"},\" The tools give you the average of what they've seen because nothing in the input asks them not to.\"]},{type:\"image\",slot:\"image2\",caption:\"Two takes on the same Patina brief. Left, an output one designer called “the clearest articulation of Patina's concept… the kind of design that makes you slow down rather than scroll past.” Right, one that drew the opposite verdict: “the typography is inconsistent, the border radiuses are inconsistent, the spacing and shadows are chaotic… it looks so template-ish.”\"},{type:\"p\",text:\"AI coding tools can produce design. The question is whether a designer would put their name on what comes back. We ran 24 outputs from Cursor, Claude Code, Codex, and Antigravity past five working designers in three rounds: blind tournament, individual ratings, then the one that decides everything. Would you present this to a client?\"},{type:\"p\",text:\"The answer split.\"},{type:\"image\",slot:\"image3\",caption:\"The split in one card. Cursor takes the tournament; Claude Code takes the client meeting.\"},{type:\"p\",text:\"Cursor won the head-to-head tournament with about a 60% overall pairwise win rate and the highest Elo rating at 1575. In a side-by-side comparison, designers picked Cursor more often than any other tool. Claude Code came second at ~50%, Antigravity third at ~45%, and Codex last at ~40%.\"},{type:\"image\",slot:\"image4\",caption:\"Overall pairwise win rate across head-to-head matchups per tool.\"},{type:\"p\",text:\"But when designers were asked whether they'd present the output to a client, Claude Code led at 63%. Antigravity followed at 53%. Cursor, the pairwise winner, dropped to 47%. And Codex landed at 30%. Designers declined to present it seven times out of ten.\"},{type:\"p\",text:\"The tool designers preferred to look at and the tool they'd trust with their name on it turned out to be different.\"},{type:\"image\",slot:\"image5\",caption:\"Client readiness flips the leaderboard. Claude Code drew the most yeses (63%), Cursor split evenly (47%), and designers declined Codex seven times in ten.\"},{type:\"h2\",text:\"What wins side-by-side vs. what wins the room\"},{type:\"p\",text:\"The gap between pairwise preference and client readiness tells you something about how designers actually evaluate work. In a tournament, they're comparing surfaces: which layout feels tighter, which color palette reads better, which typography has more presence. Client readiness seems to be an entirely different filter, with designers weighing whether they can defend every decision in the output, and if there's anything that would make them lose credibility.\"},{type:\"image\",slot:\"image6\",caption:\"Vaultra, the crypto-culture landing brief. Designers said this looked “bold and modern… it looks less like a typical crypto site” while also listing fixes: “I'd update the layout, typography, some of the colors and add icons; I don't like the arrow-style in the CTA.”\"},{type:\"p\",text:\"When designers explained why they ranked an output first, the reasons were dominated by usability and visual appeal. When they explained why they'd present something to a client, the reasons shifted toward layout, hierarchy, and whether the output felt like a real product rather than a template.\"},{type:\"p\",text:\"Designers scored outputs lower when they followed the brief but looked dated, and forgave trendy ones that drifted from it. Prompt adherence only got you in the door: outputs scoring below 2.5 on following the brief were called client-ready just 16% of the time. But once an output cleared that floor, how current it looked drove client readiness and perceived maturity more than fidelity to the brief did.\"},{type:\"h2\",text:\"Landing pages, mobile apps, web apps. Three different contests\"},{type:\"p\",text:\"The overall rankings don't convey the whole story, which is that no tool is reliably good. On a per-brief basis, Claude Code won 14 of 15 matchups on Mise (a mobile wardrobe app) and 12 of 15 on Lumen (a web analytics dashboard), but lost all 15 of its matchups on Dial (a mobile coffee-logging app), where not a single evaluator ranked it first, with a dark-mode call that left most text failing basic readability. Antigravity produced the exact mirror image, topping Dial while no evaluator ranked it first on Mise. Every output makes a single aesthetic bet, and when that bet is wrong for the use case, the whole thing collapses no matter how strong the tool is on average. Cursor held steady on most briefs without ever crashing, which is why it leads overall.\"},{type:\"image\",slot:\"image7\",caption:\"Mobile, best case against worst case. Left, a wardrobe output praised as “calm and minimal… close to the Mise brand direction.” Right, a coffee log where “most text fails basic readability standards.” The designer questioned whether dark mode was even the right call.\"},{type:\"p\",text:\"That volatility shows up again when you group by design type. On landing pages, Cursor led with a composite score of 3.65 out of 5. The composite is the average of its four rated dimensions: prompt adherence, usability, visual appeal, and modernity. On web applications, Antigravity took the lead at 3.52. Claude Code scored 3.4 on web apps, close behind. On mobile apps, everything dropped. Cursor led at 2.83, but no tool cleared 3.0.\"},{type:\"image\",slot:\"image8\",caption:\"The rankings reshuffle by format. Cursor leads landing pages (3.65) and mobile (2.83). Antigravity tops web apps (3.52).\"},{type:\"p\",text:\"Mobile was the weakest format across the board. Designers flagged the same problems regardless of tool: inconsistent spacing, navigation that didn't feel native, card layouts that packed too much information into too little space. One designer summarized their reaction to a mobile output by saying it looked like it was built by a template. Consistent padding and spacing problems that wouldn't survive a client presentation.\"},{type:\"h2\",text:\"What kept outputs from clients\"},{type:\"p\",text:\"The reasons for rejecting outputs were more uniform than the reasons for ranking them first: outputs were generic, and layouts needed structural work.\"},{type:\"p\",text:\"“Generic / needs refinement” was the most cited reason across all four tools when designers said they wouldn't present to a client, with Codex drawing 19 mentions, nearly double any other tool. But even the top performers weren't immune. Designers said yes to Claude Code's client readiness 63% of the time, but the “yes” rationales still frequently included caveats about refinement.\"},{type:\"p\",text:\"Layout / hierarchy was the second most common rejection theme. Codex drew 15 mentions, Claude Code drew 9. The complaints were specific: oversized headings wasting viewport space, filter bars that didn't look interactive, hero sections with no product below the fold. One designer described a marketplace output that was entirely a hero with nothing after it. A headline, body text, and then the page ended.\"},{type:\"image\",slot:\"image9\",caption:\"Same Mise brief, two outcomes. Left drew structural complaints: “the top and bottom menus do not feel stable… the information hierarchy needs improvement.” Right won the room: “it focuses on the user's demand… it communicates the brand identity… filters work and it's fast.”\"},{type:\"p\",text:\"Mobile apps drew the toughest feedback. Evaluators pointed to issues with information architecture, visual styles that felt outdated, and outputs not refined enough for the brief. Of the three format types, mobile had the widest gap between what designers expected and what the tools delivered.\"},{type:\"p\",text:\"Designers also classified each output by how far along it felt. Across all four tools, production-ready was the exception: no tool got more than a quarter of its outputs there, and most landed at brainstorming or mockup. Codex barely reached production-ready at all.\"},{type:\"image\",slot:\"image10\",caption:\"Design-stage classification per tool. Production-ready was the exception; most outputs landed at brainstorming or mockup.\"},{type:\"h2\",text:\"The verdict\"},{type:\"p\",text:\"Every tool won something and every tool lost something. Cursor took the most head-to-head matchups but split the room on client readiness. Claude Code drew the most client-ready verdicts but collapsed on briefs that didn't suit it. Antigravity swung between the best output in the set and the worst. Codex trailed on nearly every measure.\"},{type:\"p\",text:\"The finding that matters most is that client readiness and side-by-side preference run on different logic. In the tournament, all four dimensions we rated predicted winning a matchup at similar strength, with prompt adherence marginally the strongest. Client readiness flipped that. Once an output met a baseline of adherence, what mattered was whether it felt current. Whether a designer looked at it and thought, this feels like something shipping today. Prompt adherence wins the side-by-side. Taste wins the client's yes. And the tools that won did so by understanding the product. The ones that lost often produced visually capable work that simply missed what was asked for.\"},{type:\"p\",text:\"So the advice for a designer picking these up today is specific: reach for Cursor on landing pages, Antigravity or Claude Code on web apps, and for mobile, pick whichever you know best and budget real rework since no tool cleared the bar. And treat whatever comes back as a starting point. The tool gives you an opening move. Finishing the work is still the designer's job.\"},{type:\"p\",text:\"The gap between “looks good in a side-by-side” and “I'd put my name on this” is the gap these tools still need to close.\"},{type:\"h2\",text:\"Methodology\"},{type:\"p\",text:\"Five working designers in UI/UX, web, and mobile app design sourced from Contra's top-earning talent evaluated six briefs across three format types: two landing pages, two mobile apps, and two web applications, each set in a distinct product domain (crypto, developer tooling, specialty coffee, analytics, luxury fashion, and curated e-commerce). Briefs were written to professional standards and contained a project overview, objective, target audience, brand identity, mood references, and required screens or sections. No color palettes or typography specs were provided, forcing each tool to make its own aesthetic decisions from the brief's tone and references alone.\"},{type:\"p\",text:\"Each brief was run through all four tools using each tool's default model configuration, producing 24 total outputs:\"},{type:\"list\",ordered:false,items:[\"Cursor — cursor-agent v2026.05.24-dda726e, composer-2.5\",\"Claude Code — v2.1.150, claude-sonnet-4-6\",\"OpenAI Codex — v0.134.0, gpt-5.5-medium\",\"Google Antigravity — v1.0.2, gemini-3.5-flash-high\"]},{type:\"p\",text:\"A single standardized prompt was used per brief across all tools:\"},{type:\"code\",language:\"text\",text:\"You are an expert web developer and visual designer. Build a complete, polished web page based on the user's instructions.\\n\\nTechnical requirements:\\n- Generate a complete and valid HTML document with DOCTYPE, meta charset, and viewport meta tag.\\n- Return raw HTML that can be used directly without any additional processing.\\n- NO frameworks or libraries (React, Vue, Angular, Tailwind, Bootstrap, etc.). Pure vanilla HTML, CSS, and JavaScript only.\\n- When an external dependency is genuinely needed (e.g., an icon set or animation library), load it from UNPKG via a script/link tag.\\n- Use semantic HTML elements (nav, main, section, article, header, footer, etc.).\\n- All interactive elements must have visible hover/focus/active states.\\n- Be accessible: use ARIA labels, sufficient color contrast (WCAG AA), focus-visible styles, and logical tab order.\\n\\nDesign requirements:\\n- Always set an explicit background color on html and body.\\n- Define a cohesive color palette using CSS custom properties on :root.\\n- Establish a clear typographic hierarchy. Import a web font from Google Fonts if it elevates the design.\\n- Use generous whitespace and a consistent spacing rhythm.\\n- Apply subtle visual polish: box-shadows, border-radius, smooth transitions on interactive states.\\n\\nCSS architecture:\\n- Write all styles in a single