PinchBench Benchmark Skill

PinchBench measures how well LLM models perform as the brain of an OpenClaw agent. Results are collected on a public leaderboard at pinchbench.com.

Prerequisites

• Python 3.10+
• uv package manager
• OpenClaw instance (this agent)

Quick Start

cd <skill_directory>

# Run benchmark with a specific model
uv run benchmark.py --model anthropic/claude-sonnet-4

# Run only automated tasks (faster)
uv run benchmark.py --model anthropic/claude-sonnet-4 --suite automated-only

# Run specific tasks
uv run benchmark.py --model anthropic/claude-sonnet-4 --suite task_01_calendar,task_02_stock

# Skip uploading results
uv run benchmark.py --model anthropic/claude-sonnet-4 --no-upload

Available Tasks (23)

Command Line Options

Token Registration

To submit results to the leaderboard:

# Register for an API token (one-time)
uv run benchmark.py --register

# Run benchmark (auto-uploads with token)
uv run benchmark.py --model anthropic/claude-sonnet-4

Results

Results are saved as JSON in the output directory:

# View task scores
jq '.tasks[] | {task_id, score: .grading.mean}' results/0001_anthropic-claude-sonnet-4.json

# Show failed tasks
jq '.tasks[] | select(.grading.mean < 0.5)' results/*.json

# Calculate overall score
jq '{average: ([.tasks[].grading.mean] | add / length)}' results/*.json

Adding Custom Tasks

Create a markdown file in tasks/ following TASK_TEMPLATE.md. Each task needs:

• YAML frontmatter (id, name, category, grading_type, timeout)
• Prompt section
• Expected behavior
• Grading criteria
• Automated checks (Python grading function)

Leaderboard

View results at pinchbench.com. The leaderboard shows:

• Model rankings by overall score
• Per-task breakdowns
• Historical performance trends