OpenAI strengthens GPT-Rosalind, its model dedicated to life sciences

OpenAI upgrades GPT-Rosalind, its life sciences model, outperforming GPT-5.5, Grok 4.3 and Gemini 3.1 Pro on LifeSciBench with up to 31% fewer tokens.

Designed for enterprise-scale life sciences research, GPT-Rosalind is based on a new version that combines the agentic coding and tool-use capabilities of GPT-5.5 with enhanced mastery of medicinal chemistry and genomics.

OpenAI evaluates it via LifeSciBench, a benchmark assessed by external experts and structured into six stages of scientific work, from evidence processing to results communication, where the model outperforms GPT-5.5, Grok 4.3, and Gemini 3.1 Pro. Gains are also observed in more targeted evaluations in medicinal chemistry and genomics, as well as in laboratory benchwork assistance, where the model improves from 56% to 63%, all with reduced token consumption, up to 31% less in genomics.

Beyond reasoning, two plugins, Life Sciences Research and NGS Analysis, add an execution layer within Codex: retrieval of sourced evidence, bioinformatics analyses, and provenance tracking, complemented by native visualizers for sequences, alignments, and structures. Deployment is through a restricted research preview, via a trusted access framework, to organizations conducting public benefit work with robust governance and security, with Novo Nordisk among the first partners. OpenAI is also extending these capabilities to the field of public health and biodefense, through its Rosalind Biodefense program.