O
#1 OverallGPT-4.1
OpenAI
Highest overall score with strong balance across all categories.
- Overall
- 82.6
- Factuality
- 88.7
- Speed
- 78.4
Watch for: Can over-structure simple creative work.
AI model benchmarks for real growth work
CresciBench evaluates leading LLMs on the work growth teams actually ship: creative, SEO, localization, factuality, speed, and cost.
8
models
30
tasks
6
categories
Jun 16, 2026
updated
Public leaderboard
Best all-around model for real growth work. Scores are normalized from 0 to 100. Higher is better except raw cost, which is shown in USD per 1M tokens.
Pilot benchmark snapshot - Jun 16, 2026
| Rank | Model | Overall ↓ | Factuality | Speed | Cost | Best for |
|---|---|---|---|---|---|---|
11 | O GPT-4.1 OpenAI | 82.6 | 88.7 | 78.4 | $2.10 | All-around performance |
21 | A Claude 3.7 Sonnet Anthropic | 79.8 | 86.1 | 62.3 | $3.00 | Reasoning and long context |
3- | G Gemini 1.5 Pro | 76.4 | 83.2 | 70.1 | $1.60 | Research and analysis |
4- | M Llama 3.1 70B Instruct Meta | 72.1 | 78.3 | 93.7 | $0.59 | Cost-efficient scale |
51 | M Mistral Large 2 Mistral AI | 68.7 | 74.6 | 86.2 | $0.95 | Speed and efficiency |
61 | C Cohere Command R+ Cohere | 66.3 | 73.1 | 61.4 | $1.20 | Enterprise use cases |
7- | P Phi-3.5 Mini Instruct Microsoft | 61.2 | 66.9 | 123.5 | $0.20 | Low cost operations |
8- | G Gemma 2 27B | 58.4 | 62.7 | 82.9 | $0.30 | Open models and customization |
Scores are pilot values for the static public snapshot. Live runner data comes next.
View all modelsModel profiles
Compact readouts for the models that define the pilot set.
A
#1 FactualityClaude 3.7 Sonnet
Anthropic
Best factual accuracy and strongest long-form reasoning behavior.
- Overall
- 79.8
- Factuality
- 86.1
- Speed
- 62.3
Watch for: Slower and more expensive on short operator tasks.
M
#1 SpeedLlama 3.1 70B Instruct
Meta
Fastest generation among leading open-weight models in the pilot set.
- Overall
- 72.1
- Factuality
- 78.3
- Speed
- 93.7
Watch for: Needs stricter evaluation on factual tasks.
P
#1 CostPhi-3.5 Mini Instruct
Microsoft
Lowest cost per token with acceptable performance on simple tasks.
- Overall
- 61.2
- Factuality
- 66.9
- Speed
- 123.5
Watch for: Not strong enough for final strategic output.
Benchmark tasks
A preview of the task suite. Each prompt is designed around operator work instead of toy examples.
View all 30 tasks
1SEO - Blog outline
SEOCategory avg: 78.4
SEO - Blog outline
SEOPrompt
Create a detailed blog outline about AI content strategies for B2B SaaS for a growth marketer.
Best model output (GPT-4.1)
- Define the buying-stage search intent before the article structure.
- Separate operator workflow, content governance, and measurement sections.
- Include internal-link targets and FAQ blocks for conversion support.
2Localization - Product description
LocalizationCategory avg: 75.2
Localization - Product description
LocalizationPrompt
Adapt a growth-service product description into English, Italian, and Spanish while preserving conversion intent.
Best model output (Claude 3.7 Sonnet)
- Keeps the offer concrete instead of translating word-for-word.
- Changes CTA pressure by market without losing the core promise.
- Flags idioms that should not be reused across languages.
3Creative - Ad copy
CreativeCategory avg: 71.6
Creative - Ad copy
CreativePrompt
Write five UGC ad hooks and a 20-second script for a founder selling an Instagram growth service.
Best model output (GPT-4.1)
- Leads with a specific founder pain rather than a generic growth promise.
- Includes visual beats, proof placement, and a low-friction CTA.
- Avoids engagement-bait language that would weaken trust.
4Hallucination - Factual consistency
HallucinationCategory avg: 82.9
Hallucination - Factual consistency
HallucinationPrompt
Summarize a provider comparison from a short evidence packet and refuse claims not present in the source text.
Best model output (Claude 3.7 Sonnet)
- Separates supported claims from uncertain inferences.
- Declines to invent pricing or uptime numbers.
- Keeps confidence labels tied to the supplied evidence.
Methodology
Built to become transparent and reproducible as the live runner comes online.
Real growth tasks
The pilot suite covers content, SEO, localization, creative direction, and factuality checks.
Multi-dimensional scoring
Models are evaluated on usefulness, factuality, speed, cost, and operator readiness.
Transparent inputs
Prompts, rubric notes, output previews, and benchmark dates stay visible on the page.
Same conditions
Runs are designed to use the same prompt, temperature, and output limits per task.
Versioned snapshots
Public data is shown as an approved snapshot so visitors never see half-finished runs.