OpenClaw

Open-source agentic framework with deep shell integration, recursive planning, and multi-provider support.

91.8-0.4% (24h)

Non-Gameable Scoring

Scores are derived from established benchmarks, adjusted for harness-specific performance across four dimensions: Coding, Reasoning, Tool Use, and Autonomy.

Each dimension starts from public benchmark data and applies harness-specific modifiers based on tool integration, context handling, and orchestration quality. The overall score is a weighted composite that penalizes narrow optimization.

Model	Overall	Coding	Reasoning	Tool Use	Autonomy
Claude Sonnet 4.6	93.5	95	93	94	91
GPT-5.4	91.2	92	94	89	88
Grok 3	90.1	89	91	90	87