Score vs. cost
Average task cost vs overall score
Each dot is one model. X axis is average cost per benchmark task, including model and judge cost; Y axis is average calibrated score.
Back to evals
business benchmark collection
Landing Pages
Benchmarks for testing whether models can turn a product brief into a clear, persuasive, conversion-aware landing page.
Which models can create landing pages that are clear, specific, persuasive, and buildable?
4 benchmarks
12 tests
276 completed runs
20 base models
At a glance
Top model
claude-opus-4.8-high
83.08
Lowest cost / eval
qwen3.5-plus-02-15
$0.0113
Median rank score
79.67
Last refresh
2026-06-02
Overall ranking
Top models by average score
Higher is better. Scores come from completed judged runs.
Benchmark heatmap
Model performance by benchmark
Cells are colored by rank within each benchmark: the top ten are split across greens, anything below the top ten is red.
Below top 10
#1
| Rank | Model | Overall | Hero Clarity Test | Five-Second Clarity Test | Objection Handling Test | Landing Page Structure Test |
|---|---|---|---|---|---|---|
| 1 |
claude-opus-4.8-high
12 scored tests |
83.1 | 82.3 | 83.0 | 85.3 | 81.7 |
| 2 |
kimi-k2.5
12 scored tests |
82.8 | 81.0 | 79.7 | 85.0 | 85.3 |
| 3 |
claude-opus-4.6-high
12 scored tests |
82.5 | 85.0 | 83.7 | 80.0 | 81.3 |
| 4 |
claude-opus-4.8-low
12 scored tests |
82.4 | 82.7 | 79.0 | 84.7 | 83.3 |
| 5 |
qwen3.7-max
12 scored tests |
82.2 | 80.7 | 82.0 | 85.0 | 81.0 |
| 6 |
claude-opus-4.7
12 scored tests |
81.4 | 80.3 | 79.0 | 85.3 | 81.0 |
| 7 |
claude-opus-4.8
12 scored tests |
81.3 | 78.7 | 80.0 | 85.3 | 81.3 |
| 8 |
gemini-3-flash-preview
12 scored tests |
81.2 | 80.7 | 79.0 | 84.0 | 81.0 |
| 9 |
qwen3.5-plus-02-15
12 scored tests |
80.8 | 77.3 | 82.0 | 82.0 | 81.7 |
| 10 |
gpt-5.5-pro
12 scored tests |
80.7 | 78.7 | 81.3 | 81.3 | 81.3 |
| 11 |
claude-opus-4.6
12 scored tests |
80.3 | 81.7 | 80.7 | 81.0 | 78.0 |
| 12 |
gpt-5.4-mini
12 scored tests |
79.7 | 77.7 | 77.3 | 83.3 | 80.3 |
| 13 |
gpt-5.5
12 scored tests |
79.7 | 75.7 | 79.3 | 82.7 | 81.0 |
| 14 |
glm-5.1
12 scored tests |
79.3 | 77.3 | 80.0 | 83.7 | 76.3 |
| 15 |
gpt-5.4
12 scored tests |
78.7 | 82.3 | 77.3 | 78.0 | 77.0 |
| 16 |
deepseek-v3.2
12 scored tests |
78.2 | 75.3 | 78.7 | 79.7 | 79.0 |
| 17 |
gpt-5.4-nano
12 scored tests |
77.5 | 77.7 | 79.3 | 72.3 | 80.7 |
| 18 |
claude-sonnet-4.6
12 scored tests |
76.8 | 77.0 | 79.0 | 79.3 | 71.7 |
| 19 |
gemini-3.5-flash-high
12 scored tests |
76.1 | 77.7 | 80.3 | 84.3 | 62.0 |
| 20 |
glm-5
12 scored tests |
75.6 | 75.0 | 79.0 | 81.3 | 67.0 |
| 21 |
grok-4.20-beta
12 scored tests |
75.1 | 72.7 | 78.3 | 68.7 | 80.7 |
| 22 |
gemini-3.1-pro-preview
12 scored tests |
74.5 | 78.7 | 82.0 | 85.0 | 52.3 |
| 23 |
minimax-m2.7
12 scored tests |
67.8 | 72.7 | 76.7 | 69.3 | 52.7 |
Full leaderboard
Quality, cost, and speed
| Model | Score | Tests | Avg cost / task | Avg seconds / task | Frequent problems |
|---|---|---|---|---|---|
|
claude-opus-4.8-high
|
83.08 Strong | 12/12 | $0.0349 | 24.1s | - |
|
kimi-k2.5
|
82.75 Strong | 12/12 | $0.0130 | 64.0s | Unsupported invention |
|
claude-opus-4.6-high
|
82.5 Strong | 12/12 | $0.0365 | 37.0s | Incomplete output Unsupported invention |
|
claude-opus-4.8-low
|
82.42 Strong | 12/12 | $0.0361 | 24.6s | Wrapper text |
|
qwen3.7-max
|
82.17 Strong | 12/12 | $0.0123 | 57.6s | - |
|
claude-opus-4.7
|
81.42 Strong | 12/12 | $0.0394 | 31.5s | Incomplete output Unsupported invention Wrapper text |
|
claude-opus-4.8
|
81.33 Strong | 12/12 | $0.0340 | 22.7s | Wrapper text |
|
gemini-3-flash-preview
|
81.17 Strong | 12/12 | $0.0168 | 18.1s | Unsupported invention |
|
qwen3.5-plus-02-15
|
80.75 Strong | 12/12 | $0.0113 | 56.3s | - |
|
gpt-5.5-pro
|
80.67 Strong | 12/12 | $0.1658 | 49.3s | - |
|
claude-opus-4.6
|
80.33 Strong | 12/12 | $0.0361 | 36.7s | Incomplete output Unsupported invention |
|
gpt-5.4-mini
|
79.67 Usable | 12/12 | $0.0146 | 14.0s | - |
|
gpt-5.5
|
79.67 Usable | 12/12 | $0.0309 | 22.9s | - |
|
glm-5.1
|
79.33 Usable | 12/12 | $0.0120 | 64.2s | Incomplete output |
|
gpt-5.4
|
78.67 Usable | 12/12 | $0.0220 | 21.3s | Wrapper text |
|
deepseek-v3.2
|
78.17 Usable | 12/12 | $0.0122 | 32.9s | - |
|
gpt-5.4-nano
|
77.5 Usable | 12/12 | $0.0129 | 14.4s | Wrapper text Missing required element |
|
claude-sonnet-4.6
|
76.75 Usable | 12/12 | $0.0287 | 32.7s | Incomplete output Unsupported invention Wrapper text |
|
gemini-3.5-flash-high
|
76.08 Usable | 12/12 | $0.0286 | 20.2s | Incomplete output Missing required element |
|
glm-5
|
75.58 Usable | 12/12 | $0.0124 | 48.5s | Wrapper text Incomplete output Missing required element |
|
grok-4.20-beta
|
75.08 Usable | 12/12 | $0.0138 | 17.3s | Unsupported invention Wrapper text |
|
gemini-3.1-pro-preview
|
74.5 Usable | 12/12 | $0.0336 | 27.2s | Incomplete output |
|
minimax-m2.7
|
67.83 Needs editing | 12/12 | $0.0124 | 36.9s | Unsupported invention Incomplete output Missing required element |
Test cases
Where the scores come from
Each row is one prompt, with score distributions, top and low performers, and the most frequent problems judges flagged.
| Test | Benchmark | Avg | Max | Min | Top model | Lowest model | Frequent problems |
|---|---|---|---|---|---|---|---|
|
AI bookkeeping tool for freelancers landing_hero_001 |
Hero Clarity Test | 76.8 | 86.0 | 70.0 | claude-opus-4.6-high · 86 | minimax-m2.7 · 70 | Wrapper text ×2 |
|
Compliance tracker for clinics landing_hero_002 |
Hero Clarity Test | 78.9 | 84.0 | 73.0 | kimi-k2.5 · 84 | qwen3.5-plus-02-15 · 73 | Wrapper text ×2 |
|
Hiring platform for hospitality teams landing_hero_003 |
Hero Clarity Test | 80.2 | 86.0 | 74.0 | claude-opus-4.7 · 86 | claude-opus-4.8 · 74 | Wrapper text ×3 Unsupported invention ×1 |
|
Internal knowledge bot for law firms landing_5sec_001 |
Five-Second Clarity Test | 81.4 | 86.0 | 74.0 | claude-opus-4.6-high · 86 | claude-opus-4.8 · 74 | Wrapper text ×1 |
|
AI survey analysis tool landing_5sec_002 |
Five-Second Clarity Test | 81.0 | 85.0 | 74.0 | claude-opus-4.7 · 85 | claude-opus-4.8-low · 74 | Wrapper text ×1 |
|
Warehouse shift planning software landing_5sec_003 |
Five-Second Clarity Test | 77.2 | 82.0 | 68.0 | gemini-3.1-pro-preview · 82 | claude-opus-4.7 · 68 | Unsupported invention ×7 |
|
AI customer support agent objections landing_objection_001 |
Objection Handling Test | 80.1 | 86.0 | 65.0 | claude-opus-4.8 · 86 | minimax-m2.7 · 65 | Unsupported invention ×2 Missing required element ×1 |
|
Payroll software migration landing_objection_002 |
Objection Handling Test | 79.9 | 85.0 | 60.0 | gpt-5.4-mini · 85 | minimax-m2.7 · 60 | Unsupported invention ×4 Wrapper text ×2 |
|
Automated outbound sales tool landing_objection_003 |
Objection Handling Test | 83.5 | 86.0 | 76.0 | claude-opus-4.7 · 86 | gpt-5.4-nano · 76 | - |
|
B2B analytics tool page structure landing_structure_001 |
Landing Page Structure Test | 73.2 | 85.0 | 32.0 | gpt-5.5 · 85 | gemini-3.5-flash-high · 32 | Incomplete output ×7 Missing required element ×3 Wrapper text ×1 |
|
AI writing platform for law firms landing_structure_002 |
Landing Page Structure Test | 82.3 | 86.0 | 70.0 | claude-opus-4.7 · 86 | claude-sonnet-4.6 · 70 | Incomplete output ×3 |
|
Personal finance app for expats landing_structure_003 |
Landing Page Structure Test | 73.7 | 86.0 | 40.0 | kimi-k2.5 · 86 | minimax-m2.7 · 40 | Incomplete output ×6 Wrapper text ×3 |
Model profiles
Strengths, weaknesses, and tradeoffs
claude-opus-4.8-high
12 scored tests · Strong
83.08
Highest traits
credibility8.6
risk handling8.6
tone8.5
objection coverage8.5
audience fit8.38
Lowest traits
differentiation7.9
proof placement8.0
cta quality8.03
buildability8.13
structure8.17
kimi-k2.5
12 scored tests · Strong
82.75
Highest traits
risk handling8.53
objection coverage8.5
structure8.47
buildability8.47
proof placement8.47
Lowest traits
trust7.73
cta quality7.82
differentiation7.97
message clarity8.2
clarity8.23
claude-opus-4.6-high
12 scored tests · Strong
82.5
Highest traits
audience fit8.67
clarity8.5
differentiation8.43
conversion logic8.4
specificity8.39
Lowest traits
credibility7.7
objection coverage8.1
cta quality8.12
structure8.17
buildability8.17
claude-opus-4.8-low
12 scored tests · Strong
82.42
Highest traits
credibility8.5
risk handling8.5
objection coverage8.5
structure8.43
audience fit8.43
Lowest traits
cta quality8.0
trust8.07
differentiation8.07
message clarity8.23
buildability8.3
qwen3.7-max
12 scored tests · Strong
82.17
Highest traits
credibility8.5
risk handling8.5
objection coverage8.47
tone8.37
specificity8.28
Lowest traits
differentiation7.5
cta quality7.85
structure8.03
conversion logic8.07
buildability8.1
claude-opus-4.7
12 scored tests · Strong
81.42
Highest traits
credibility8.6
risk handling8.6
tone8.5
objection coverage8.5
audience fit8.47
Lowest traits
trust6.93
cta quality7.95
buildability8.0
differentiation8.03
structure8.1
claude-opus-4.8
12 scored tests · Strong
81.33
Highest traits
credibility8.7
risk handling8.7
tone8.6
objection coverage8.5
audience fit8.37
Lowest traits
differentiation7.83
cta quality7.9
structure8.03
proof placement8.07
conversion logic8.13
gemini-3-flash-preview
12 scored tests · Strong
81.17
Highest traits
credibility8.5
risk handling8.47
objection coverage8.33
audience fit8.28
tone8.23
Lowest traits
differentiation7.57
trust7.67
cta quality7.83
structure8.0
buildability8.07
qwen3.5-plus-02-15
12 scored tests · Strong
80.75
Highest traits
credibility8.3
risk handling8.3
objection coverage8.23
message clarity8.2
buildability8.17
Lowest traits
differentiation6.73
cta quality7.83
specificity8.05
tone8.1
audience fit8.12
gpt-5.5-pro
12 scored tests · Strong
80.67
Highest traits
risk handling8.27
credibility8.2
message clarity8.2
buildability8.17
conversion logic8.17
Lowest traits
differentiation7.4
cta quality7.78
tone8.0
trust8.0
structure8.03
claude-opus-4.6
12 scored tests · Strong
80.33
Highest traits
tone8.5
audience fit8.4
clarity8.4
objection coverage8.33
proof placement8.27
Lowest traits
structure7.5
trust7.7
buildability7.73
differentiation7.73
credibility7.83
gpt-5.4-mini
12 scored tests · Usable
79.67
Highest traits
credibility8.5
risk handling8.5
objection coverage8.33
tone8.17
buildability8.13
Lowest traits
cta quality7.33
differentiation7.33
audience fit7.85
specificity7.86
trust7.9
Compare models
Trait profile and test-by-test matchup
Pick two models to compare average trait scores and individual test outcomes from the published calibrated results.
Per-test comparison
Sorted by the biggest absolute score difference between the selected models.
| Test | Benchmark | Delta | ||
|---|---|---|---|---|
Frequent problems
Where models break
Incomplete output
16
Wrapper text
15
Unsupported invention
14
Missing required element
4
Example outputs
Winner, middle, failure
winner
kimi-k2.5 · 86
The model demonstrates unusually deep task-specific judgment by incorporating highly accurate niche details (FBAR, FATCA, data residency, repatriation) into a sophisticated, highly buildable conversion flow.
**1. The Expat Reality Check (Hero)** * **Title:** "Your Money Lives in Three Countries. Your Dashboard Should Too." * **Purpose:** Immediate pattern interruption + identification of the core struggle; establish the "multi-jurisdictional" pain point within 3 seconds. * **Key Copy Points:** * Headline acknowledges the fragmentation (accounts in London, taxes in Berlin, investments in Singapore). * Subhead: "One dashboard to track real-time currency exposure, cross-border tax liabilities, and account balances—without the 3 AM spreadsheet panic." * Primary CTA: "Create your expat money map" ( prominent, high-contrast button). * Risk reversal: "Free for 30 days. No credit card required to map your first country." * **Visual/Component:** Interactive world map visualization showing animated data lines connecting 3-4 major financial hubs (e.g., NYC, London, Tokyo) converging into a single, clean mobile interface. Auto-plays on loop to demonstrate consolidation. **2. The Hidden Tax Trap (Problem Agitation)** * **Title:** "Are You Accidentally Tax Non-Compliant?" * **Purpose:** Agitate the specific anxiety of expat tax complexity (FBAR, FATCA, double taxation treaties) to heighten urgency. * **Key Copy Points:** * "You track expenses in euros, earn in dollars, and owe taxes in both." * "Missing one FBAR filing can cost $10,000. We auto-flag reporting thresholds across your jurisdictions." * "Currency swings just ate your rent budget—again." * **Visual/Component:** Split-screen "Anxiety Dashboard" vs. "Clarity Dashboard." Left side: chaotic spreadsheet screenshots, red alert icons, and overlapping calendar reminders. Right side: Clean, green-status tax checklist with countdown timers to filing deadlines per country. **3. The Money Map Method (Solution)** * **Titl
failure
gemini-3.5-flash-high · 32
The model's response cuts off abruptly after the second section, failing to provide a complete landing page structure, missing the final CTA, and failing to fully address the constraints.
Here is a conversion-aware landing page structure tailored for B2B SaaS growth and product teams. --- ### Section 1: The Hero * **Title:** Stop Guessing Why Your Free Trials Don't Convert * **Purpose:** Instantly grab the attention of growth and product leaders by stating the exact problem solved, the primary benefit, and a low-friction call to action. * **Key Copy Points:** * **Headline:** Turn trial drop-offs into paying customers. * **Subheadline:** Map every friction point in your trial user journey automatically. See exactly where users get stuck, fix the leaks, and boost your trial-to-paid conversion rate by an average of 18%. * **Primary CTA Button:** Request a Demo * **Micro-copy (under CTA):** *See your trial bottlenecks in a 15-minute personalized walkthrough.* * **Suggested Visual/Component:** A high-fidelity, interactive mock-up of the dashboard. It should show a user funnel with a highlighted drop-off point (e.g., a red warning circle over the step: *"Invited team member - 72% drop-off"*). --- ### Section 2: Integration & Speed Bar * **Title:** Plugs Directly Into Your Existing Growth Stack * **Purpose:** Build immediate technical credibility and handle the "how much dev work is this?" objection right out of the gate. * **Key Copy Points:** * No new tracking scripts required. We sync with the tools you already use. * Setup is usually complete in under 2 weeks. * One-click integrations with: **Segment**, **HubSpot**, **Stripe**, and **PostHog**. * **Suggested Visual/Component:** A clean, horizontal logo bar featuring Segment, HubSpot, Stripe, and PostHog, with a subtle "Connected" green check
mid performer
deepseek-v3.2 · 81
The model perfectly follows all constraints, delivering clear, concise, and highly relevant copy that directly addresses the target audience's objections with the provided proof points.
**Headline:** Your Firm's Knowledge. Your Private AI. Zero Hallucinations. **Subheadline:** A secure internal AI assistant that answers questions directly from your approved documents—and cites every source. Deployed inside your firm's private cloud, never on public servers. **Benefit Bullets:** - **Confidentiality Assured:** Operates entirely within your firm’s private cloud—no external data exposure. - **Answers You Can Trust:** Every response is grounded in your internal documents and includes verifiable source citations. - **Instant Expertise:** Quickly find precise information across memos, case files, and policies without manual searches. **Primary CTA:** Book a Security Review **Secondary CTA:** See a Confidential Demo
Methodology
How scores are produced
Each model output is scored by a strict JSON LLM judge, supported by deterministic heuristics, then normalized to a 0-100 score.
LLM judge
A stable judge model scores each trait from 0 to 10 using benchmark-specific traits.
Heuristics
Deterministic checks catch length, banned phrases, required sections, format validity, and safety flags.
Calibrated ceiling
Rubrics are intentionally strict, leaving room above today's best scores for future model improvements.