ClawBench (OpenClaw) - Dev: leaderboard

ClawBench (2026, NAIL Group) tests web agents on everyday write-heavy online tasks such as purchases, bookings and applications on live websites, stopping the final state-changing request and judging each run against a human reference trajectory. This board is the task success rate in percent on the developer and technology tasks with agents in the OpenClaw harness; higher is better.

Metric: Success rate (%) on the 18 developer and tech tasks of the 153-task ClawBench paper evaluation (agents in the OpenClaw browser harness on live production websites, the final state-changing request intercepted before it reaches the server, each run judged pass or fail by an Agent-as-Judge (Claude Sonnet 4.6) against a human reference trajectory); higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Haiku 4.527.8#583
2Gemini 3.1 Pro (Preview)22.2#99
3Gemini 3 Flash22.2#175
4Qwen 3.5 (ClawBench checkpoint unspecified)22.2
5GLM-516.7#269
6GPT-5.4High11.1#135
7Claude Sonnet 4.6High11.1#173
8Gemini 3.1 Flash Lite5.6#371

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=clawbench-openclaw-dev · How It Works · Data refreshed daily, snapshot 2026-11-21.