OpenAI GPT-Rosalind: Drug Discovery AI Deployed at Amgen

OpenAI's newest model doesn't write code or summarize emails. It reads scientific papers, queries genomics databases, and designs molecular cloning protocols — and it already runs inside Amgen and Moderna. GPT-Rosalind, OpenAI's first dedicated life sciences AI, targets one of medicine's most stubborn bottlenecks: the 10-to-15-year journey from a promising drug target to an FDA-approved treatment. That timeline is about to shrink — for the labs that can get access.

The 15-Year Drug Discovery Problem GPT-Rosalind Was Built to Shrink

Drug discovery is notoriously slow and brutally expensive. A pharmaceutical company identifies a promising drug target, then spends a decade-plus running experiments, failing, redesigning, and navigating regulatory approval — all before a single patient sees the treatment. Most of that time isn't spent on creative scientific breakthroughs. It's spent on analytical drudgework: reading thousands of research papers, designing experimental protocols, cross-referencing genomic databases, and writing evidence summaries that any PhD scientist could produce — if they had an entire year to do nothing else.

GPT-Rosalind is built to automate exactly that drudgework. Think of it as a senior research associate that has read every paper ever published on your molecule of interest, and can cross-reference genomics, biochemistry, and clinical literature simultaneously — in minutes rather than months.

The model's core capabilities include:

• Evidence synthesis — parsing thousands of published papers and surfacing relevant findings automatically
• Hypothesis generation — proposing experimental directions based on existing data patterns
• Experimental planning — designing end-to-end molecular protocols, including complete cloning (the process of copying a specific DNA segment and inserting it into a host cell for study) procedures
• Database integration — connecting to 50+ specialized scientific tools and datasets via the Life Sciences plugin (a software add-on that links the model to external research databases) for Codex
• Multi-step research tasks — handling the full arc from literature review to experimental pathway suggestion in a single interface, without switching tools

GPT-Rosalind Benchmark Results: What Pharma R&D Is Paying Attention To

95th Percentile Against Human Experts — on Unpublished Data

OpenAI tested GPT-Rosalind on RNA (ribonucleic acid — the molecule that translates genetic instructions into proteins) sequence-to-function prediction, evaluated by Dyno Therapeutics on novel, unpublished biological data. This detail matters: the model couldn't have memorized the answers during training. It had to reason from scratch, competing directly against human scientists submitting their own work. GPT-Rosalind's best-of-ten submissions ranked at the 95th percentile among human expert submissions. On sequence generation tasks using equally novel unpublished data, it ranked at the 84th percentile.

These aren't synthetic lab conditions — they're evaluations on real research questions that real scientists were simultaneously trying to answer.

Outperforming GPT-5.4 Across Specialized Scientific Tasks

On LABBench2 (a standardized suite of 11 scientific reasoning tasks used to benchmark life sciences AI performance across disciplines), GPT-Rosalind outperformed GPT-5.4 — OpenAI's most capable general-purpose frontier model — on 6 out of 11 tasks. The clearest advantage appeared in CloningQA, which tests a model's ability to design complete molecular cloning protocols from scratch. On BixBench, a bioinformatics (the use of software to analyze large biological datasets such as DNA sequences and gene expression patterns) benchmark measuring data analysis performance, GPT-Rosalind recorded a 0.751 pass rate.

For a domain-specialized model to outperform a frontier general-purpose model on the majority of domain-specific tasks is significant. It validates the argument that the next phase of AI value in science won't come from making models bigger — it will come from making them more precisely tuned to specific scientific fields.

Named After the Scientist Whose Credit Was Stolen

The model's name isn't marketing filler. Rosalind Franklin (1920–1958) was a British chemist and X-ray crystallographer (a scientist who uses X-ray beams to map the atomic-level structure of molecules) whose painstaking laboratory work produced Photo 51 — the crucial X-ray image that revealed DNA's double-helix structure. James Watson and Francis Crick used her data without her permission or knowledge. When the Nobel Prize was awarded in 1962 for the discovery of DNA's structure, Franklin was not included. She had died four years earlier from ovarian cancer at age 37.

Her foundational imaging work underpins virtually every modern genomics tool — including every AI model now attempting to decode genetic sequences. Naming GPT-Rosalind after her is OpenAI's clearest symbolic acknowledgment yet of a scientist who shaped modern biology and received almost none of the recognition for it.

Who's Running GPT-Rosalind Today — and What They're Building

GPT-Rosalind is not publicly available. Access is gated — restricted to US-based qualified enterprise customers enrolled in OpenAI's trusted-access program. The four confirmed early partners represent life sciences at institutional scale:

• Amgen — Fortune-500 biotech developing treatments for cancer, cardiovascular disease, and inflammatory conditions, with over $33 billion in annual revenue
• Moderna — the mRNA vaccine pioneer now applying mRNA technology (a method of delivering genetic instructions to cells, originally used to produce COVID-19 vaccines) to cancer therapies and rare diseases
• Allen Institute — a nonprofit research institute founded by Microsoft co-founder Paul Allen, specializing in neuroscience and cell biology
• Thermo Fisher Scientific — a $40+ billion analytical instruments and lab services company that supplies the physical infrastructure for global pharmaceutical R&D

Beyond pharmaceutical applications, OpenAI is collaborating with Los Alamos National Laboratory (the US government research facility founded during the Manhattan Project, now focused on national security science and energy research) on AI-guided protein and catalyst design — work that reaches beyond drug discovery into materials science and clean energy applications.

These are not startups experimenting with AI. These are the organizations that design the drugs prescribed to millions of people and supply the equipment that makes clinical research possible. Their early adoption — and willingness to participate in a gated access program — signals genuine institutional confidence in the model's utility.

The Access Wall — and the Real Reason It Exists

For most researchers — individual scientists, academic labs at smaller institutions, international teams — GPT-Rosalind is effectively inaccessible today. The model is:

• US-only — international life sciences researchers cannot enroll regardless of institutional affiliation or credentials
• Enterprise-gated — the trusted-access program is designed for large institutional partners, not individual researchers or early-stage startups
• Closed-source — no model weights, no fine-tuning access, no self-hosted deployment option

The access restrictions aren't purely commercial strategy. GPT-Rosalind includes technical safeguards designed to detect and flag potentially dangerous use cases — specifically, requests that could inform the synthesis of hazardous biological agents. It's a genuine tradeoff: some legitimate edge-case scientific workflows may be blocked by automated safety flags, adding friction that open-source alternatives wouldn't impose. OpenAI appears to have decided that controlled deployment with institutional accountability is worth the slower rollout.

Their own framing is deliberate: "The model is definitely not intended to replace scientists, but rather to help them move faster through some of the most time-intensive and analytically demanding stages of the scientific process." GPT-Rosalind is positioned as a force multiplier for scientists who can access it — not an autonomous replacement for domain expertise.

What Drug Discovery Timeline Compression Means for Patients Who Can't Wait

The practical stakes are high. If GPT-Rosalind genuinely compresses drug development timelines by even 2-3 years — eliminating the analytical drudgework that researchers already know slows them down — the downstream impact is measured in lives. A cancer drug that reaches patients in 11 years instead of 14 means three more years of treatment availability for people diagnosed today. An Alzheimer's therapy that exits clinical trials faster means millions of families don't wait as long. Every analytical step that gets automated is time returned to people who don't have it to spare.

OpenAI has signaled that GPT-Rosalind is a strategic direction, not a one-off experiment. The shift toward domain-specific AI models — each benchmarked, named, and targeted at a specific scientific problem — is the company's answer to Google DeepMind's protein folding dominance and Anthropic's growing presence in scientific reasoning research. This is version one. Watch for what comes next.

If you work in pharmaceutical R&D, genomics research, or biotech and are US-based, you can contact OpenAI's enterprise team for trusted-access program enrollment. For a broader look at how AI automation is transforming research workflows across industries, explore our guides section.

Related Content — Get Started with AI Tools | Guides | More AI News