OpenAI introduced BioBasicsBench, a 100-task benchmark for AI agents in wet-lab science. It tests interpretation of raw experimental data like gels, microscopy and instrument readouts. Tasks cover protein methods, imaging, assays, nucleic acids and cloning. Scientists from Ligo co-developed the benchmark using real lab data and expert grading rubrics. GPT-6 Astra scored 66.7% at high reasoning effort, up from 43.8% for GPT-5.6 Sol. Astra improved at recognizing data limits, handling multimodal inputs and retaining experimental context. Models still struggle with assumptions overriding controls and ignoring contradictory evidence. The benchmark aims to push AI toward reliable lab assistance.
What changed
Existing benchmarks used curated charts and images from scientific papers, often missing real-world ambiguity and troubleshooting needs.
What it unlocks
Measuring AI agents' ability to interpret raw wet-lab data and handle everyday experimental ambiguity.
- 100 tasks in BioBasicsBench
- 66.7% mean reward for GPT-6 Astra at xhigh
- 43.8% mean reward for GPT-5.6 Sol at xhigh
- 18,500 total evaluation results
Sources