🦞

PinchBench

Claw-some AI Agent Testing

PinchBench

Find the best model for your OpenClaw

Run the benchmark yourself →

Include unofficial runsOpen-weight onlyUpdated 03/25/2026, 4:27 PM

50 models608 runs

Include unofficial runsOpen-weight onlyUpdated 03/25/2026, 4:27 PM

🦀

Success rate by model

💰 Budget filter:Max $per run

Percentage of tasks completed successfully across standardized OpenClaw agent tests

Scores are graded via automated checks and LLM judge. How we benchmark·View all tasks

Best ScoreAverage Score

🦞anthropic/claude-opus-4.6

🦀openai/gpt-5.4

🦐qwen/qwen3.5-27b

qwen/qwen3.5-397b-a17b

minimax/minimax-m2.7

xiaomi/mimo-v2-flash

nvidia/nemotron-3-super-120b-a12b

anthropic/claude-sonnet-4.5

minimax/minimax-m2.1

anthropic/claude-sonnet-4.6

Totally An Ad

Hosted OpenClaw — your personal AI agent, managed by Kilo.

Hosting and inference cost for PinchBench sponsored by Kilo, so we totally hope you try KiloClaw so we can keep the lights on around here.

From $8/month + AI inference at cost

Model	Provider	Best %Best %	Avg %
🦞`anthropic/claude-opus-4.6`		93.3%93.3%	82.9%
🦀`openai/gpt-5.4`		90.5%90.5%	81.8%
🦐`qwen/qwen3.5-27b`		90.0%90.0%	78.5%
`qwen/qwen3.5-397b-a17b`		89.1%89.1%	80.4%
`minimax/minimax-m2.7`		89.0%89.0%	82.5%
`xiaomi/mimo-v2-flash`		88.8%88.8%	70.2%
`nvidia/nemotron-3-super-120b-a12b`		88.6%88.6%	76.7%
`anthropic/claude-sonnet-4.5`		88.6%88.6%	81.1%
`minimax/minimax-m2.1`		88.4%88.4%	80.2%
`anthropic/claude-sonnet-4.6`		88.0%88.0%	80.8%

All tasks and grading criteria are open source. Hover column headers for details.