PinchBench Leaderboard

Fetches and formats the PinchBench leaderboard — AI agent benchmarks for LLMs on standardized OpenClaw coding tasks.

Workflow

1. Determine the query

Map the user's intent to script flags:

User intent	Flags
"Show the leaderboard" / default	--top 10
"Top 5 models"	--top 5
"How does Claude perform?"	--model claude
"Cheapest models"	--sort cost --top 10
"Fastest models"	--sort time --top 10
"Compare Gemini and Claude"	Run twice with --model gemini and --model claude, present side by side
"Full leaderboard"	--top 50

2. Run the script

{
  "tool": "exec",
  "command": "python3 {baseDir}/scripts/fetch_leaderboard.py --top 10"
}

Available flags:

• --top N — number of models to show (default: 10)
• --sort metric — sort by score, cost, time, or runs (default: score)
• --model filter — filter models containing this string (case-insensitive)
• --json — output raw JSON for further processing

3. Format the response

Present the output as-is in a code block. Add a brief one-line insight after the table:

• Highlight the top performer and its score
• If the user asked about a specific model, comment on its ranking relative to the field
• If sorting by cost, note the best value (score/cost ratio)

4. Error handling

• If the script fails with a curl error → report the error, suggest checking network connectivity
• If the script fails to parse data → the site structure may have changed, inform the user
• If no models match the filter → say so and suggest a broader search

Examples

User says	Flags	Expected behavior
"Show me the PinchBench leaderboard"	--top 10	Show top 10 by score
"Which model is cheapest for OpenClaw?"	--sort cost --top 10	Show top 10 sorted by cost
"How does Claude compare to GPT?"	--model claude then --model gpt	Show both, compare
"What's the fastest model on PinchBench?"	--sort time --top 5	Show top 5 by execution time