A frontier model tuned for the lab bench
OpenAI pushed a significant update to GPT-Rosalind, its model purpose-built for life-sciences research at enterprise scale. The pitch is that it now combines GPT-5.5's agentic coding and tool use with deeper domain intelligence across drug discovery - medicinal chemistry, genomics, quantitative biology, and wet-lab troubleshooting.
The benchmarks - and the token-efficiency angle
OpenAI built a set of expert-judged evaluations and reports gains that come with better efficiency, which matters as labs watch their compute bills:
- On MedChemBench (structure-activity reasoning, ADME prediction, retrosynthesis and more), it edged GPT-5.5 while using about 7% fewer tokens.
- On GeneBench, a long-horizon agentic genomics evaluation, it improved accuracy while using 31% fewer tokens.
- On LabWorkBench, built on proprietary and therefore uncontaminated wet-lab protocols, it posted a clear jump over GPT-5.5 on real troubleshooting and optimization tasks.
- A broader benchmark, LifeSciBench, scores end-to-end work across six workflow areas rather than testing a single skill in isolation.
From reasoning to executed workflows
The more practical story is execution. Two new plugins - Life Sciences Research and Life Sciences NGS Analysis - pull evidence retrieval, biological interpretation, and bioinformatics execution into one workspace while preserving provenance. Both are available to all users through Codex, with qualified enterprise users able to power them with GPT-Rosalind itself. OpenAI also added interactive viewers for native biological file types - sequence, alignment, and structure - so a scientist can inspect the evidence in-context as the model reasons across a workflow; the demo walks through a tumor-biopsy analysis that narrows to a specific mutation and pulls in resistance context.
Access, partners, and the safety framing
Access is expanding to eligible organizations globally through a trusted-access deployment model reserved for groups doing legitimate research with clear public benefit, strong governance, and enterprise-grade security. Novo Nordisk is an early named partner, using the model to connect evidence across literature, genomics, and experimental results to speed drug-discovery decisions. OpenAI is also offering a managed workspace for qualified organizations that lack an Enterprise account. The company frames the whole effort as advancing biological capability alongside safeguards - explicitly tying it to its parallel Rosalind Biodefense work.