Real-world benchmarks for AI coding agents
Note: This repository contains the benchmark skill/tasks. It is NOT the source of official leaderboard results. To add models to the official results, modify pinchbench/scripts/default-models.yml.
PinchBench measures how well LLM models perform as the brain of an OpenClaw agent. Instead of synthetic tests, we throw real tasks at agents: scheduling meetings, writing code, triaging email, researching topics, and managing files.
Results are collected on a public leaderboard at pinchbench.com.
Most LLM benchmarks test isolated capabilities. PinchBench tests what actually matters for coding agents:
- Tool usage โ Can the model call the right tools with the right parameters?
- Multi-step reasoning โ Can it chain together actions to complete complex tasks?
- Real-world messiness โ Can it handle ambiguous instructions and incomplete information?
- Practical outcomes โ Did it actually create the file, send the email, or schedule the meeting?
# Clone the skill
git clone https://github.com/pinchbench/skill.git
cd skill
# Run benchmarks with your model of choice
./scripts/run.sh --model openrouter/anthropic/claude-sonnet-4
# Or run specific tasks
./scripts/run.sh --model openrouter/openai/gpt-4o --suite task_calendar,task_stockNote: Model IDs must include their provider prefix (e.g.
openrouter/,anthropic/). OpenRouter is the default provider used for routing.
Requirements:
- Python 3.10+
- uv package manager
- A running OpenClaw instance
| Category | Tasks | What's tested |
|---|---|---|
| Productivity | Calendar, daily summaries | Event creation, time parsing, scheduling |
| Research | Stock prices, conferences, markets | Web search, data extraction, synthesis |
| Writing | Blog posts, emails, humanization | Content generation, tone, formatting |
| Coding | Weather scripts, file structures | Code generation, file operations |
| Analysis | Spreadsheets, PDFs, documents | Data processing, summarization |
| Triage, search | Inbox management, filtering | |
| Memory | Context retrieval, knowledge management | Long-term memory, recall |
| Skills | ClawHub, skill discovery | OpenClaw ecosystem integration |
Each task is graded automatically, by an LLM judge, or both โ ensuring both objective and nuanced evaluation.
To get your results on the leaderboard:
# Register for an API token (one-time)
./scripts/run.sh --register
# Run benchmark โ results auto-upload with your token
./scripts/run.sh --model openrouter/anthropic/claude-sonnet-4Skip uploading with --no-upload if you just want local results.
To submit an official run (marked on the leaderboard):
# Using environment variable
export PINCHBENCH_OFFICIAL_KEY=your_official_key
./scripts/run.sh --model anthropic/claude-sonnet-4
# Using command line flag
./scripts/run.sh --model anthropic/claude-sonnet-4 --official-key your_official_key| Flag | Description |
|---|---|
--model MODEL |
Model to test (e.g., openrouter/anthropic/claude-sonnet-4) |
--judge MODEL |
Judge model for LLM grading; uses direct API when set (see below) |
--suite SUITE |
all, automated-only, or comma-separated task IDs |
--runs N |
Number of runs per task for averaging |
--timeout-multiplier N |
Scale timeouts for slower models |
--output-dir DIR |
Where to save results (default: results/) |
--no-upload |
Skip uploading to leaderboard |
--register |
Request an API token for submissions |
--upload FILE |
Upload a previous results JSON |
--official-key KEY |
Mark submission as official (or use PINCHBENCH_OFFICIAL_KEY env var) |
By default (no --judge flag), the LLM judge runs as an OpenClaw agent session. When --judge is specified, it calls the model API directly instead, bypassing OpenClaw personality injection.
# Default: OpenClaw agent session (no --judge needed)
./scripts/run.sh --model openrouter/anthropic/claude-sonnet-4
# Direct API via OpenRouter
./scripts/run.sh --model openai/gpt-4o --judge openrouter/anthropic/claude-sonnet-4-5
# Direct API via Anthropic
./scripts/run.sh --model openai/gpt-4o --judge anthropic/claude-sonnet-4-5-20250514
# Direct API via OpenAI
./scripts/run.sh --model openai/gpt-4o --judge openai/gpt-4o
# Headless Claude CLI
./scripts/run.sh --model openai/gpt-4o --judge claudeRequired env vars: OPENROUTER_API_KEY, ANTHROPIC_API_KEY, or OPENAI_API_KEY depending on the judge model prefix.
We welcome new tasks! Check out tasks/TASK_TEMPLATE.md for the format. Good tasks are:
- Real-world โ Something an actual user would ask an agent to do
- Measurable โ Clear success criteria that can be graded
- Reproducible โ Same task should produce consistent grading
- Challenging โ Tests agent capabilities, not just LLM knowledge
Session transcripts are automatically saved to results/{run_id}_transcripts/ alongside the results JSON. Each task's full agent conversation is preserved as a JSONL file (e.g. task_calendar.jsonl) for post-run analysis.
- Leaderboard: pinchbench.com
- OpenClaw: github.com/openclaw/openclaw
- Issues: github.com/pinchbench/skill/issues
MIT โ see LICENSE for details.
Claw-some AI agent testing ๐ฆ