OpenAI introduces BioBasicsBench, a benchmark of everyday wet-lab science tasks

OpenAI introduces BioBasicsBench, a benchmark of 100 routine wet-lab tasks built with Ligo, a biotech company focused on enzyme design. The tasks use raw experimental evidence (gel images, colony plates, microscopy, sequencing files, instrument exports and videos of bench work) across protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Scientists wrote them from their own experiments, and answers are graded against expert-written rubrics that give partial credit.

OpenAI evaluated seven of its models (GPT-5.2 to GPT-6 Astra) across 37 model and reasoning-effort settings. At the xhigh setting, these are its figures for mean reward:

• GPT-6 Astra: 66.7%
• GPT-6.1 Sol: 62.0%
• GPT-5.6 Sol: 43.8%
• GPT-6 Luna: 38.3%
• GPT-5.5: 35.4%

OpenAI says Astra's gains come from stronger native multimodal handling, such as spotting that a control plate carried the wrong antibiotic label, and from recognizing when the data cannot support a conclusion. It also lists where Astra still fails: letting an assumption override the controls, discounting contradicting observations, and overcounting colonies on a plate by about 12%.