OpenAI Ships LifeSciBench and Demos a Near-Autonomous AI Chemist That Improved a Real Drug Reaction

OpenAI·high signal

On June 17 OpenAI released LifeSciBench, an expert-authored, expert-reviewed benchmark for real-world life-science research tasks, alongside a demonstration where a near-autonomous AI chemist built on GPT-5.4 (with Molecule.one) improved a challenging medicinal-chemistry reaction. The pairing signals OpenAI's push from chat into verifiable scientific R&D with a measurement harness attached. Builders should note the explicit framing that benchmark scores must be validated against real research outcomes, not taken as discovery acceleration on their own.