OpenAI Enters Drug Discovery With GPT-Rosalind, a Specialized Life Sciences Model That Ranked Above the 95th Percentile of Human Experts in RNA Prediction

Overview

OpenAI on April 16 launched GPT-Rosalind, its first AI model purpose-built for biological research and drug discovery. Named after Rosalind Franklin — the chemist whose X-ray crystallography work was pivotal to determining the structure of DNA — the model marks OpenAI’s entry into a competitive space that Google DeepMind, NVIDIA, and a cohort of AI-native biotechs have been cultivating for years.

The launch directly answers a question The Machine Herald raised in earlier coverage of Novo Nordisk’s OpenAI deal: whether the partnership would eventually involve custom models tailored to pharmaceutical workflows, rather than general-purpose models. The answer arrived two days after that partnership was announced.

What GPT-Rosalind Does

Unlike general-purpose language models, GPT-Rosalind is optimized for multi-step scientific workflows spanning chemistry, genomics, and protein engineering. According to OpenAI, the model supports four primary functions: synthesizing evidence from biomedical literature, generating hypotheses about molecular targets, planning experimental protocols, and executing multi-step research tasks across traditionally siloed disciplines.

The practical infrastructure behind the model is a Life Sciences research plugin for Codex — OpenAI’s cloud-based AI engineering agent — that connects researchers to more than 50 scientific tools and databases spanning human genetics, functional genomics, protein structure, biochemistry, and clinical evidence registries, as reported by The Next Web. The integration allows researchers to interrogate multiple data sources through a single interface rather than switching between specialized tools.

Moderna’s CEO Stéphane Bancel offered an early endorsement, stating that “GPT-Rosalind represents an important step in helping scientific teams use advanced AI to reason across complex biological evidence, data, and workflows. At Moderna, we are already seeing how it can synthesise complex data and translate those insights into experimental workflows, with the potential to accelerate the pace of R&D,” according to Euronews. Amgen, Moderna, the Allen Institute, Thermo Fisher Scientific, and Novo Nordisk are the launch partners working with OpenAI to apply the technology across their discovery processes, as reported by Euronews.

Benchmark Performance and Real-World Testing

OpenAI released two sets of performance data alongside the launch.

On BixBench — a benchmark designed around tasks that practicing bioinformaticians actually perform — GPT-Rosalind achieved the highest score among models with published results, according to The Next Web, recording a pass rate of 0.751. On LABBench2, which evaluates a broader range of laboratory reasoning tasks, the model outperformed GPT-5.4 on six of eleven sub-tasks, with the strongest gains on CloningQA, a section focused on molecular cloning protocol design.

The more significant result came from a real-world evaluation conducted with Dyno Therapeutics, a gene therapy company that provided unpublished RNA sequences the model had never seen. GPT-Rosalind’s best-of-ten submissions ranked above the 95th percentile of human experts on sequence-to-function prediction, and around the 84th percentile on sequence generation for AAV gene therapy applications, according to The Next Web. Using held-out, proprietary data for evaluation is a methodological step up from most AI benchmark releases in the life sciences.

OpenAI was careful to scope what the model can realistically achieve. The company stated it does not yet believe AI can create new disease treatments on its own. The realistic value proposition, by OpenAI’s own framing, is time compression at the hypothesis generation and experimental design stage — not clinical success prediction.

Access and Safety Controls

GPT-Rosalind is currently available as a research preview in ChatGPT, Codex, and the API, restricted to a trusted-access program for vetted enterprise customers in the United States, according to The Next Web. Organizations must demonstrate they are working toward improving human health outcomes and maintain strong security and governance controls. During the preview period, usage does not draw from existing API credits.

Los Alamos National Laboratory is also listed among the launch partners, according to The Next Web. OpenAI has built in technical safeguards to flag potentially dangerous activity, reflecting the dual-use risks inherent in genomics and biology research.

Competitive Landscape

OpenAI enters this market as a latecomer with strong workflow integration but a shorter track record of physics-grounded scientific reasoning.

Google DeepMind holds the scientific credibility advantage by a wide margin. AlphaFold2’s solution to protein structure prediction earned recognition with a Nobel Prize in Chemistry in 2024, and Isomorphic Labs — DeepMind’s drug discovery spinout — has active pharmaceutical collaborations already running. NVIDIA occupies a complementary position through its BioNeMo platform, providing GPU infrastructure and foundation models to biopharma research organizations.

AI-native biotechs like Recursion Pharmaceuticals and Schrödinger face a more complicated dynamic. Their proprietary datasets remain defensible moats, but the arrival of a general-purpose competitor that integrates with the same scientific tools and databases has introduced new competitive pressure. The core question for investors is whether OpenAI’s general-purpose approach narrows the moats of specialized firms whose value proposition is purpose-built AI for drug discovery — though specialized firms hold years of proprietary data and domain expertise that no general model can replicate quickly.

Amazon launched Amazon Bio Discovery on April 14, according to AWS, a lab-in-the-loop platform combining 40+ biological foundation models with direct integration into contract research organizations for wet-lab validation. The near-simultaneous launches by OpenAI and Amazon signal that the cloud layer is now in active competition for life sciences R&D workflows, mirroring the broader AI infrastructure race.

Analysis: The Timeline Problem

The drug discovery sector’s defining challenge is not benchmark performance — it is clinical validation over decade-long timescales. No fully AI-discovered drug has cleared Phase III trials. The most advanced AI-designed drug candidates, from firms like Insilico Medicine and Recursion, remain in early or mid-stage clinical testing.

GPT-Rosalind’s realistic contribution, by the evidence available at launch, is at the earliest stages of the discovery pipeline: identifying which experiments are worth running, synthesizing relevant literature, and surfacing molecular candidates worth investigating further. Those contributions are genuinely useful — target identification and lead optimization often consume years of researcher time before a molecule reaches clinical testing — but they are the beginning of a long road, not a shortcut to approval.

The meaningful metric to watch is not benchmark pass rates but IND filings from Amgen, Moderna, and Novo Nordisk that document GPT-Rosalind’s role in their discovery processes, followed by whether those candidates advance through development faster than comparable controls. That data will take years to accumulate. It is also the only data that ultimately matters.

What We Don’t Know

• Whether and when GPT-Rosalind will become available outside the United States, and under what governance frameworks.
• The specific financial terms of OpenAI’s partnerships with Amgen, Moderna, and Thermo Fisher Scientific.
• How access restrictions will evolve as the research preview matures into general availability.
• Whether Los Alamos National Laboratory’s involvement extends into biosecurity or dual-use research domains beyond standard health research.
• How OpenAI’s approach to biological safety review will scale as the trusted-access program expands.