OpenAI Introduces BioBasicsBench for Wet-Lab Agents

BioBasicsBench is a meaningful shift from clean academic questions toward the messy evidence scientists actually handle, though a 66.7% score still leaves agents far from dependable lab assistance.

• –Covers gels, microscopy, sequencing, instrument exports, colony plates, protocols, and bench videos across five wet-lab domains
• –Uses expert-reviewed rubrics with partial credit for conclusions, measurements, reasoning, and recognizing insufficient evidence
• –GPT-6 Astra’s 22.9-point lead over GPT-5.6 Sol suggests multimodal capability matters more than simply generating longer answers
• –The benchmark exposes practical failure modes around controls, image interpretation, calibration, and experimental context
• –Its real-world task construction makes it a useful complement to knowledge-heavy biology benchmarks