OpenAI and Ligo Jointly Release Wet Lab Daily Task Benchmark BioBasicsBench

On November 6, 2026, OpenAI and Ligo released BioBasicsBench, which evaluates the interpretation and troubleshooting capabilities of AI agents using 100 routine wet-lab tasks and raw data such as gels, petri dishes, and microscopic images. The evaluation covers 7 models and 18,500 results. Under the "xhigh" setting, the GPT-6 Astra achieved an average score of 66.7% of the maximum possible score (95% CI: 60.3–73.1%), while GPT-5.6 Sol scored 43.8%. The announcement also highlighted improvements in control interpretation, multimodal details, and experimental context. Furthermore, it noted that higher scores do not necessarily equate to reliable laboratory judgment and linked capability advancements with biosecurity governance.