OpenAI's GPT-Rosalind Upgrades Drug Discovery With 31% Fewer Tokens
GPT-Rosalind gets GPT-5.5's agentic coding and deeper domain intelligence for drug discovery, genomics, and wet lab workflows, now open to global enterprise research organizations.
Takeaways
- GPT-Rosalind gets GPT-5.5's agentic coding and tool-use, plus deeper intelligence in medicinal chemistry, genomics, and wet lab reasoning.
- Three new benchmarks show gains over GPT-5.5: MedChemBench (27.5% vs 25.1%), GeneBench (21.6% vs 20.4%), LabWorkBench (63.2% vs 55.8%) -- all with fewer tokens.
- Two open-source plugins now available via Codex: Life Sciences Research and NGS Analysis, enabling end-to-end bioinformatics workflows.
- Access expanded globally to eligible organizations via a trusted-access program; Novo Nordisk named as flagship partner alongside Amgen, Moderna, and others.
- LifeSciBench, a new externally expert-judged benchmark covering six end-to-end research workflow categories, is being opened to third-party leaderboards.
- No public pricing; organizations must request access and pass a qualification and safety review.
OpenAI's most specialized model just got a significant upgrade. GPT-Rosalind, the company's purpose-built life sciences model, has received a major capability update that grafts GPT-5.5's agentic coding and tool-use onto deeper domain intelligence in medicinal chemistry, genomics, and wet lab reasoning. This is not a general-purpose model wearing a lab coat -- it is OpenAI's first serious attempt at a vertical AI system, and the update marks a meaningful step forward in what that system can actually do.
The model is named after Rosalind Franklin, the researcher whose work was essential to understanding the molecular structures of DNA, RNA, and viruses. OpenAI shipped a model designed not to chat, generate images, or write code -- but to help discover drugs. The June update is the second major release in the series, and it raises the ceiling on what the model can execute autonomously inside a scientific workflow.
Not a chatbot, a research partner
GPT-Rosalind is OpenAI's most capable model for life sciences research, designed for deep scientific reasoning, stronger tool and database use, and safe access to multi-step workflows across areas such as protein understanding, genomics analysis, and biochemistry reasoning. The model is best suited for bioinformaticians, computational biologists, and early discovery biologists.
The distinction from a general-purpose model matters more than it might seem. A general model can explain what CRISPR is. GPT-Rosalind can reason about a guide RNA design and its predicted off-target activity. GPT-Rosalind's training skews heavily toward scientific literature, genomic databases, molecular property datasets, and clinical research records.
What the update actually adds
The update combines GPT-5.5's agentic coding and tool-use capabilities with stronger model intelligence in core drug-discovery domains such as medicinal chemistry and genomics, while advancing performance across broader life sciences analysis, design, and experimental workflows. The key word here is agentic -- meaning the model can now plan, write code, call tools, and iterate across multi-step tasks without constant human intervention.
Three new benchmarks quantify the gains:
- MedChemBench -- covers multimodal chemical structure understanding, structure-activity relationships (how a molecule's shape affects its biological activity), drug potency, toxicity, and ADME (absorption, distribution, metabolism, excretion) prediction. GPT-Rosalind scores 27.5% against GPT-5.5's 25.1%, using 7.2% fewer tokens.
- GeneBench -- an agentic evaluation on long-horizon genomics tasks spanning functional genomics, spatial transcriptomics, proteomics, epigenomics, and applied genetics. The updated GPT-Rosalind uses 31% fewer tokens than GPT-5.5 while achieving a higher accuracy of 21.6% vs. 20.4%.
- LabWorkBench -- tests the model's ability to link experimental changes to outcomes in real wet lab protocols, for troubleshooting and optimization. GPT-Rosalind scores 63.2% vs. GPT-5.5 at 55.8%, while using 5.3% fewer tokens.
The token efficiency gains across all three benchmarks are not cosmetic. In enterprise deployments running thousands of research queries, a 31% reduction in token usage translates directly to cost and latency savings at scale.

A benchmark built for real science
To anchor these evaluations in something meaningful, OpenAI introduced LifeSciBench -- an externally expert-judged benchmark designed to reflect how scientific work actually gets done. Unlike existing benchmarks that evaluate a single component of model performance or biological domain in isolation, LifeSciBench takes an end-to-end view of scientifically valuable work by drawing tasks from six workflow areas: evidence handling, analysis, design and optimization, scientific reasoning, validation and operations, and translation and communication.
A model that scores well on a narrow chemistry question-answering task is not necessarily the same model that can reconcile conflicting data from three published studies, propose an experiment, and communicate findings in regulatory-ready language. LifeSciBench attempts to capture that full arc, treating AI evaluation as a proxy for real scientific workflow performance rather than as a standardized test of domain recall. GPT-Rosalind bests other models on LifeSciBench, including GPT-5.5, Grok 4.3, and Gemini 3.1 Pro.
From reasoning to execution: the plugin layer
The intelligence upgrade is only half the story. OpenAI also shipped two open-source plugins that turn GPT-Rosalind's reasoning into repeatable, executable scientific workflows:
- Life Sciences Research plugin -- handles sourced evidence retrieval and biological interpretation, pulling from external literature and connecting it to internal data.
- Life Sciences NGS Analysis plugin -- runs bioinformatics pipelines end-to-end, including next-generation sequencing (NGS) analysis, single-cell RNA-seq QC, and bulk RNA-seq workflows.
All users can now access both plugins through Codex. Qualified GPT-Rosalind enterprise users can also use GPT-Rosalind to power the plugins. OpenAI has also added interactive viewers for biologically native file types, including sequence, alignment, and structure viewers, allowing scientists to inspect biological files while asking follow-up questions in context.
A concrete example from the announcement: a scientist investigating a liquid tumor biopsy can use the NGS Analysis plugin to surface recurring mutations and low-frequency calls, then pivot to the Research plugin to pull inhibitor and resistance context from literature, then inspect the mutant residue directly in a structure viewer -- all inside one Codex session, with full provenance preserved.

Who gets access, and how
GPT-Rosalind is now available in research preview to eligible organizations globally through OpenAI's trusted-access deployment structure. This is a meaningful expansion from the initial US-only rollout. Eligible organizations must be conducting legitimate scientific research with clear public benefit, have governance and safety oversight, and use controlled access with enterprise-grade security.
OpenAI is also offering an OpenAI-managed workspace for qualified organizations without an Enterprise account. No public pricing has been disclosed -- access is gated behind a qualification and safety review. Organizations can request access here.
The flagship named partner in this update is Novo Nordisk. The partner OpenAI spotlighted in this update is Novo Nordisk, the maker behind blockbuster GLP-1 drugs. Early partners also include Amgen, Moderna, Thermo Fisher Scientific, the Allen Institute, and Dyno Therapeutics.
The bigger picture: pharma's patent cliff and the AI race
The timing of this push is not accidental. 2026 marks the start of a particularly pronounced patent cliff, with key patents on billion-dollar drugs set to expire. The US market alone is projected to lose more than $230 billion in revenue between 2025 and 2030. Pharma companies need to refill their pipelines faster than ever, and AI-assisted discovery is one of the few levers available at that scale.
In biology, the playing field is much more open: no single model has the kind of dominant share that GPT or Claude has in code generation. That is why a credible biology launch from any frontier lab right now is a competitive event. OpenAI will face competition from Anthropic, Google DeepMind, and a growing cohort of life sciences-specific AI startups who do not apply the same access restrictions.
There is also a meaningful limitation worth flagging. A strong showing on RNA sequence-to-function prediction for AAV gene therapy says very little about whether the model can generate useful hypotheses for a small-molecule kinase inhibitor targeting a novel cancer pathway. Drug discovery is not one task -- it is hundreds of tasks, each with its own data requirements, its own failure modes, and its own standard of good enough. The benchmarks are OpenAI's own, and while LifeSciBench uses external expert graders, independent validation will be the real test.
GPT-Rosalind is a reasoning model that can plan and orchestrate scientific workflows -- including by calling structure-prediction tools like AlphaFold 3 -- and is designed to compress the cycle time between observation, hypothesis, and experiment. Whether that compression translates into approved drugs is a question that will take years to answer. But the infrastructure for a genuine AI-native drug discovery workflow is now meaningfully more complete.