OpenAI benchmarks wet-lab AI agents at 66.7% accuracy
New benchmark measures AI agent performance in wet-lab science
OpenAI released BioBasicsBench on November 6, 2026, introducing a structured evaluation framework for AI agents tasked with routine wet-lab data interpretation OpenAI. The benchmark comprises 100 concrete laboratory tasks, targeting a gap in agent evaluation: how well current models handle real-world scientific workflows that require reading instruments, interpreting protocols, and drawing conclusions from experimental data.
The benchmark directly measures agentic capability rather than abstract reasoning. BioBasicsBench focuses on everyday wet-lab workflows—the kind of tasks that research technicians and junior scientists perform daily—rather than novel discovery or experimental design. This positions the benchmark as a practical evaluation tool for enterprises and research institutions considering autonomous laboratory systems.
Performance gaps between model generations
At the "xhigh" reasoning setting, GPT-6 Astra achieved a reported mean reward of 66.7% with a 95% confidence interval of 60.3%–73.1%, averaging 4,704 output tokens per task GATE. By contrast, GPT-5.6 Sol scored 43.8% on the same benchmark, representing a significant performance gap between the two model generations.
The confidence interval and token-usage data indicate OpenAI conducted structured evaluation rather than one-off spot checks. The 23-percentage-point difference between GPT-6 Astra and GPT-5.6 Sol suggests meaningful improvements in how newer reasoning models handle multi-step laboratory inference—parsing written instructions, cross-referencing safety data, and executing conditional logic based on observed results.
Implications for laboratory automation
BioBasicsBench enters a narrow space in agent evaluation benchmarks. Most published benchmarks measure coding, web navigation, or mathematical reasoning. Laboratory science requires different competencies: attention to protocol detail, understanding of measurement uncertainty, and awareness of safety constraints. A two-thirds success rate at GPT-6 Astra's level raises practical questions about where autonomous agents can safely augment or replace human work in wet labs.
The benchmark makes agent performance in laboratory science concrete and measurable rather than speculative. Institutions evaluating whether to pilot autonomous lab systems now have a reproducible reference point for comparing model capability OpenAI. Future releases of BioBasicsBench or similar benchmarks could track whether continued model scaling and reasoning improvements push laboratory agent performance closer to production readiness.
