LABBench 2Reasoning

Real-world biology research tasks; models work in a Linux terminal with preinstalled bioinformatics tools, Python, and R, plus restricted internet access.LABBench 2 evaluates AI agents on real biology research tasks. Models work in a Linux terminal with preinstalled bioinformatics tools, Python, and R, with restricted internet access, and carry out end-to-end research workflows such as data processing, analysis, and hypothesis validation.

Ranking Sources

#ModelScore

1

Gemini 3.7 FlashGoogle

82.1%

2

GPT-5.6 TerraOpenAI

81.2%

3

Claude Sonnet 5Anthropic

80.1%

4

Gemini 3.6 FlashGoogle

76.1%

4 models total

Back to benchmarks