⏱ Capsule Public Vibe Design

Terra & Kiln — Landing Page + Design System

Ezarwebmaster· Jul 20, 2026

Brief

Rationale This capsule is not a discrimination-style Benchy test in the classic sense — it's a trend observatory. The product brief and content are fixed so that 100+ runs remain directly comparable; the only variable is the design system each model invents. Replayed over time on the Timeline, this captures shifts in aesthetic trends (e.g. "this month it's warm terracotta minimalism, in six months something else") — something no other benchmarking tool can do. The sector (small-batch ceramics studio) was chosen because it carries no dominant AI design cliché (unlike B2B SaaS, which reliably pulls toward the same purple-gradient pattern), leaving maximum room to observe genuine creative variance. The "wow factor" is intentionally left out of the GT as a pass/fail criterion — beauty is not a fact verifiable by calculator/regex, unlike the rest of the GT. The prompt gives explicit permission and incentive to take creative risks (asymmetry, micro-interactions, scroll reveals), but final aesthetic judgment stays a human signal via capsule_likes/run_ratings, not a capsule score. Expected spread Frontier models: coherent, well-documented design systems; tasteful image treatment; at least one genuinely distinctive layout/interaction moment executed cleanly Mid-tier models: solid token discipline but a "safe" generic layout (centered hero, plain 3-column cards) despite the explicit prompt nudge toward distinctiveness Weaker models: declared tokens not honored in practice (especially in hover/focus states), missing sections, or an attempt to import an external animation/icon library despite the added creative freedom Known cheats / risks <details> ignored or design system documentation placed elsewhere (e.g. back at the top) — direct structural violation Product images dropped in with zero visual treatment (no consistent filter/crop/frame) — violates "consistent visual treatment" requirement A different image path/URL referenced, or a stray external image added — direct fail, explicit constraint violated Tokens declared in :root but not respected elsewhere, especially in :hover/:focus states (the classic leak already observed on a prior run) The newly granted creative freedom (subtle JS, animations) used as a loophole to import an external animation or icon library via CDN — must be checked first given the added latitude Third-party image host (Imgur) dependency is a known, accepted temporary permanence risk — flagged for swap to Supabase Storage; the prompt itself will not need to change, only the three URLs Grading Truth (GT) #CriterionPassPartialFailGT1Renders with no console errors; only tolerated network requests are the three given image URLsokminor non-blocking errorsblocking error or request to a non-whitelisted domainGT2Native <details> present, closed by default, positioned right before the footer, opens without JScompliantposition/label slightly offdocumentation outside <details>, or JS required to openGT3Fidelity between documented values and actual computed style, applied everywhere including hover/focusnear-perfect matchminor discrepanciesdocumentation disconnected from the restGT4The 3 images used at the correct URLs, no other image/URL referencedcompliant1 image missing or misreferencedpaths ignored or external image usedGT5Consistent visual treatment across the 3 images (crop/ratio, CSS filter, framing — a clear system, not raw copy-paste)clearly unified treatmentpartial/inconsistent treatmentraw images, no treatmentGT6Required sections present, in the specified orderall present1 missing/incomplete2+ missingGT7Responsive at 375/768/1440 with no overflow or broken elementsclean at all 3 sizes1 size degradedbroken at 2+ sizesGT8Copy consistent with brand values (slowness, imperfection, durability)clearly integratedvaguely evokedgeneric copy Scoring 16 points max (8 criteria × 2 — Pass=2, Partial=1, Fail=0): S: 15–16 — flawless token discipline, zero cheats, clean creative execution A: 12–14 — solid, 1 minor weakness B: 8–11 — functional but token discipline or responsiveness imperfect C: 4–7 — missing sections or detected cheat (external image, network request, CDN import) F: 0–3 — unrenderable or major technical constraint violation Qualitative observation (non-scored) Grading reports should include a free-text note per run on the "wow factor" — is there a genuinely distinctive layout/interaction moment, or does the run default to a safe generic pattern despite the prompt's nudge? This complements capsule_likes/run_ratings as a qualitative signal for the trend-observatory purpose of this series, without affecting the tier score.

Locked Reference Prompt

IMMUTABLE

Scientific timeline lock active

You are designing a landing page for a fictional small-batch ceramics studio called "Terra & Kiln." Brand facts (fixed — use exactly as given): - Name: Terra & Kiln - One-line description: Handmade stoneware and porcelain pieces, thrown and fired in small batches by a single studio. - Values to convey somewhere in the copy: slowness, imperfection as beauty (wabi-sabi adjacent), durability, everyday use over display-only objects. Three product photos are provided and MUST be used exactly as given, at these fixed URLs: - Mug photo: https://qoyuqtiafuqqyarrnsxr.supabase.co/storage/v1/object/public/Assets/Mug.jpg - Bowl photo: https://qoyuqtiafuqqyarrnsxr.supabase.co/storage/v1/object/public/Assets/Bowl.jpg - Vase photo: https://qoyuqtiafuqqyarrnsxr.supabase.co/storage/v1/object/public/Assets/Vase.jpg Do not reference any other image URL, any other CDN, or generate/invent any additional image source. These three URLs are the only images available, and each must be used exactly once. Build a single, fully self-contained HTML file (inline CSS, inline/vanilla JS only — no external requests of any kind except the three image URLs above) containing a complete landing page with the following sections, in this exact order: 1. Hero — brand name, a short tagline, one clear call-to-action button 2. Studio story — a short "about" section conveying the brand values above 3. Product showcase — exactly 3 product cards, one for the mug photo, one for the bowl photo, one for the vase photo, each with a product name and short description. Apply a consistent visual treatment to all three images (e.g. consistent aspect ratio/crop, consistent CSS filter or color-grading, consistent border/frame treatment) so they read as part of the same designed system — do not just drop the raw images in with no treatment. 4. Newsletter / contact signup — an email input and submit button (non-functional is fine, but must be visually complete) 5. A collapsed, native HTML `<details>` element (with a `<summary>` labeled something like "View Design System") placed just before the footer, closed by default, containing your invented design system documentation: - Color palette: each color shown as a swatch with its name/role (e.g. "Primary", "Accent", "Background") and its exact hex or rgb value printed as visible text next to the swatch - Typographic scale: a sample line for each type size you use (e.g. H1, H2, H3, Body, Caption), each labeled with its font size and weight as visible text - Spacing scale: a visual representation of your spacing units (e.g. small boxes or bars at each spacing step) labeled with their pixel/rem values - Button/component states: your primary and secondary button styles shown side by side, labeled 6. Footer — basic links and copyright line Design requirements: - Invent your own complete design system: a color palette, a typographic scale, and a spacing scale. Declare these as CSS custom properties and apply them consistently across the ENTIRE page (hero, story, products, newsletter, and the design system documentation itself) — do not hardcode one-off colors, font sizes, or spacing values anywhere, including in hover/focus states. - The `<details>` element must use only native browser behavior (no JavaScript required for it to open/close). - The page must be fully responsive and look intentional at mobile (~375px), tablet (~768px), and desktop (~1440px) widths. - Do not use any external fonts, icon libraries, or CDNs. The only external requests allowed are the three image URLs given above. Beyond the structural and technical requirements above, aim for something visually memorable — the kind of page a design-conscious visitor would want to screenshot and share, not a generic, safe template. Concretely: - Include at least one distinctive, deliberate visual or interaction moment that goes beyond the baseline requirements — e.g. an unexpected hero composition, an asymmetrical or editorial layout choice, a subtle scroll-triggered reveal, or a thoughtful micro-interaction on hover/focus. Pick whatever fits your design system, not a specific technique we're prescribing. - Avoid the most generic, overused patterns (perfectly centered hero text over a flat color, three identical icon-title-text columns, a plain stacked list of cards with no visual rhythm) unless you deliberately execute them with unusual craft. - Subtle CSS transitions/animations and small amounts of vanilla JS (e.g. for scroll-triggered reveals via IntersectionObserver) are allowed, as long as there are still zero external network requests and no console errors — do not import any animation library, icon library, or script from a CDN to achieve this. Output only the complete HTML file, nothing else.

Add a Benchmark Run

Sign in to run this prompt against hundreds of models with your own OpenRouter key and archive the results.

Timeline (10 runs)

Run Activity

10 runs in the last 6 months

Jan
Feb
Mar
Apr
May
Jun
Jul
LessMore

Cost vs Speed

size = output tokens · top-left is best

100$0.010$0.10$1.00fast · cheapfast · priceyslow · cheapslow · priceyCost per request ($, log)Speed (tokens/sec, log)
size = tokens4.7k19k47k
value frontier — no model is faster & cheaper

Capsule Stats

Runs
10
Total cost
$2.2053
Tokens
228k
Avg latency
6m 12s
Models
10
Web Searches
0
Fastest nvidia/nemotron-3-ultra-550b-a55b37s
Cheapest deepseek/deepseek-v4-flash$0.0046
Top provider DeepSeek×1
100k reasoning last today
Tip: Select 2 or more runs via their "Compare" buttons — or filter by company below and compare them all at once — then open the Compare Studio: verdicts, benchmark bars, charts, side-by-side reading and response diff.

Benchmark Run — Jul 20, 2026 Latest

Moonshot AI (Kimi)
Moonshot AI (Kimi)kimi-k3
moonshotai/kimi-k3Jul 20, 2026, 07:11:01 PM
Extended
stop
Latency
19m 46s
client → response
Input Tokens
1,116
prompt tokens
Output
41,555
generated
Total Tokens
42,671
in + out
Billed Cost
$0.6267
OR Credits
Reasoning
29,039
thinking tokens
Model Output
kimi-k3 — Canvas

Benchmark Run — Jul 20, 2026

NVIDIA
NVIDIAnemotron-3-ultra-550b-a55b
nvidia/nemotron-3-ultra-550b-a55bJul 20, 2026, 04:40:41 PM
Extended
stop
Latency
37s
client → response
Input Tokens
1,072
prompt tokens
Output
11,642
generated
Total Tokens
12,714
in + out
Billed Cost
$0.0286
OR Credits
Reasoning
6,110
thinking tokens
Model Output
nemotron-3-ultra-550b-a55b — Canvas

Benchmark Run — Jul 20, 2026

KwaiPilot (Kuaishou)
KwaiPilot (Kuaishou)kat-coder-air-v2.5
kwaipilot/kat-coder-air-v2.5Jul 20, 2026, 10:41:26 AM
Extended
stop
Latency
10m 9s
client → response
Input Tokens
1,068
prompt tokens
Output
15,850
generated
Total Tokens
16,918
in + out
Billed Cost
$0.0097
OR Credits
Reasoning
1,028
thinking tokens
Model Output
kat-coder-air-v2.5 — Canvas

Benchmark Run — Jul 20, 2026

Anthropic
Anthropicclaude-fable-5
anthropic/claude-fable-5Jul 20, 2026, 04:14:02 AM
Extended
stop
Latency
2m 24s
client → response
Input Tokens
1,631
prompt tokens
Output
14,223
generated
Total Tokens
15,854
in + out
Billed Cost
$0.7275
OR Credits
Reasoning
854
thinking tokens
Model Output
claude-fable-5 — Canvas

Benchmark Run — Jul 20, 2026

Z Ai
Z Aiglm-5.2
z-ai/glm-5.2Jul 20, 2026, 04:10:55 AM
Extended
stop
Latency
12m 11s
client → response
Input Tokens
1,049
prompt tokens
Output
47,150
generated
Total Tokens
48,199
in + out
Billed Cost
$0.1456
OR Credits
Reasoning
31,628
thinking tokens
Model Output
glm-5.2 — Canvas

Benchmark Run — Jul 20, 2026

OpenAI
OpenAIgpt-5.6-sol
openai/gpt-5.6-solJul 20, 2026, 04:04:56 AM
Extended
stop
Latency
6m 35s
client → response
Input Tokens
1,033
prompt tokens
Output
11,160
generated
Total Tokens
12,193
in + out
Billed Cost
$0.3400
OR Credits
Reasoning
1,421
thinking tokens
Model Output
gpt-5.6-sol — Canvas

Benchmark Run — Jul 20, 2026

MiniMax
MiniMaxminimax-m3
minimax/minimax-m3Jul 20, 2026, 04:03:35 AM
Extended
stop
Latency
3m 45s
client → response
Input Tokens
1,207
prompt tokens
Output
29,113
generated
Total Tokens
30,320
in + out
Billed Cost
$0.0353
OR Credits
Reasoning
16,825
thinking tokens
Model Output
minimax-m3 — Canvas

Benchmark Run — Jul 20, 2026

Google DeepMind
Google DeepMindgemini-3.1-pro-preview
~google/gemini-pro-latestJul 20, 2026, 04:03:04 AM
Extended
stop
Latency
1m 47s
client → response
Input Tokens
1,086
prompt tokens
Output
17,026
generated
Total Tokens
18,112
in + out
Billed Cost
$0.2065
OR Credits
Reasoning
8,490
thinking tokens
Model Output
gemini-3.1-pro-preview — Canvas

Benchmark Run — Jul 20, 2026

xAI (Grok)
xAI (Grok)grok-4.5
x-ai/grok-4.5Jul 20, 2026, 03:49:27 AM
Extended
stop
Latency
2m 16s
client → response
Input Tokens
1,272
prompt tokens
Output
13,095
generated
Total Tokens
14,367
in + out
Billed Cost
$0.0809
OR Credits
Reasoning
396
thinking tokens
Model Output
grok-4.5 — Canvas

Benchmark Run — Jul 20, 2026

DeepSeek
DeepSeekdeepseek-v4-flash
deepseek/deepseek-v4-flashJul 20, 2026, 03:49:27 AM
Extended
stop
Latency
2m 24s
client → response
Input Tokens
1,062
prompt tokens
Output
16,011
generated
Total Tokens
17,073
in + out
Billed Cost
$0.0046
OR Credits
Reasoning
4,694
thinking tokens
Model Output
deepseek-v4-flash — Canvas