QwenClawBench
Real-world agent benchmark — 100 tasks across 8 domains. [qwenclawbench-v1.1-100]
12
Models
100
Tasks
Top Score
60.5
Claude Sonnet 4.6
Sort
Score
12 models
↵ click any row for per-task breakdown
Scoring Methodology
Three grading modes depending on what the task requires.
Auto
Automated
Python functions verify outputs directly. Binary pass/fail with optional partial credit.
LLM
LLM Judge
Claude scores the agent's output against a rubric, handling open-ended or hard-to-verify answers.
Hybrid
Hybrid
A weighted combination of automated and LLM scores. However, if the automated score falls below a minimum threshold, the LLM component is zeroed — the agent must pass the hard checks first.
Hybrid Formula
final = weight_auto × auto + weight_llm × llm
but if auto < 0.75, treat llm as 0 (hard checks must pass first)
but if auto < 0.75, treat llm as 0 (hard checks must pass first)
Passes hard checks
w=(0.6/0.4)
A:90%
L:85%
= 88.0%
Fails hard checks — LLM zeroed
w=(0.6/0.4)
A:70%
L:95%
= 42.0%
Perfect auto, weak LLM
w=(0.6/0.4)
A:100%
L:50%
= 80.0%