It takes 10 to 15 years to get a drug from lab bench to pharmacy shelf. OpenAI thinks its new AI model can help shorten that timeline. But the researchers who might benefit most from it probably won’t get access anytime soon.
GPT-Rosalind is OpenAI’s first domain-specific model, built from the ground up for biology, drug discovery, and translational medicine. Named after Rosalind Franklin — the British scientist whose X-ray crystallography work was essential to discovering the structure of DNA — the model launched last week as a research preview. It’s available through ChatGPT, Codex, and the API, but only for “qualified customers” who pass OpenAI’s vetting process.
Those customers are exactly who you’d expect: Amgen, Moderna, the Allen Institute, Thermo Fisher Scientific, and — following a separate partnership announced April 14 — Novo Nordisk.
What GPT-Rosalind Actually Does
This isn’t a chatbot that happens to know some biology. GPT-Rosalind is a frontier reasoning model optimized for multi-step scientific workflows: synthesizing research literature, analyzing genetic sequences, planning experiments, interpreting protein structures, and processing research data.
The benchmarks are genuinely impressive. On BixBench, which tests real-world bioinformatics and data analysis, GPT-Rosalind scored 0.751 — beating GPT-5.4 (0.732), Grok 4.2 (0.698), and Gemini 3.1 Pro (0.550). On LABBench2, a battery of research tasks from literature retrieval to molecular cloning protocol design, it outperformed GPT-5.4 on six of eleven tasks.
When tested through Codex, the model’s best submissions ranked above the 95th percentile of human experts on prediction tasks and around the 84th percentile on sequence generation.
Alongside the model, OpenAI released a Life Sciences research plugin for Codex — a set of modular skills connecting researchers to over 50 scientific tools and data sources, covering human genetics, functional genomics, protein structure, biochemistry, and clinical evidence.
Not AlphaFold’s Competitor — Its Orchestrator
Press coverage has positioned GPT-Rosalind as OpenAI’s challenge to Google DeepMind’s AlphaFold, but that framing misses the point. AlphaFold solved a specific problem: predicting 3D protein structure from amino acid sequences. GPT-Rosalind doesn’t replace that — it’s designed to sit on top of it.
The model works as an orchestration and reasoning layer. It takes AlphaFold’s structural outputs and integrates them with genomic data, published research, and experimental planning. Think of AlphaFold as a microscope and GPT-Rosalind as the researcher deciding what to look at and why.
That said, OpenAI is absolutely competing with DeepMind for pharmaceutical partnerships. Google’s Isomorphic Labs spinoff has already secured roughly $3 billion in deals with Eli Lilly and Novartis. OpenAI landing Novo Nordisk, Amgen, and Moderna signals it wants a piece of the same market.
The Access Problem
Here’s where things get complicated. GPT-Rosalind is U.S. enterprise only, gated behind qualification reviews and safety assessments. OpenAI calls this a “trusted access program.”
The justification is biosecurity. A model that can reason through molecular biology, design experimental protocols, and predict sequence-to-function relationships could theoretically be used to engineer biological threats as easily as therapeutic treatments. An international coalition of over 100 scientists has called for tighter controls on biological data used to train AI, citing pathogen design risks.
Those concerns are real. But critics point out that gating is a deployment control, not a capability control. The underlying model exists. The real biosecurity test comes when comparable capability lands in open-weight systems — and given the pace of open-source AI development, that’s a question of when, not if.
There’s a more immediate problem too. The organizations that got early access — Amgen, Moderna, Thermo Fisher — are massive pharmaceutical companies with existing AI infrastructure. University researchers, smaller biotech startups, and scientists in developing countries where drug-resistant diseases hit hardest are left waiting.
As one analysis put it bluntly: “OpenAI is selling governed capability — vetted users and audit trails — as much as it’s selling reasoning power.” The safety framing conveniently doubles as a business strategy that creates scarcity and drives enterprise pricing.
The Benchmark Reality Check
GPT-Rosalind’s numbers are strong. But there’s a gap between benchmark performance and clinical impact that no one should ignore.
No fully AI-discovered drug has cleared Phase 3 clinical trials to date. A model can show a better benchmark score this week, but proving that its recommendations actually changed the fate of a drug pipeline takes years. The real success metric isn’t accuracy on curated RNA tasks — it’s whether GPT-Rosalind helps researchers avoid dead-end programs that would have otherwise consumed years and billions of dollars.
OpenAI’s Joy Jiao was careful to set expectations, noting the model cannot autonomously create treatments but can “help researchers move faster through some of the most complex and time-intensive parts of the scientific process.”
That’s an honest framing. It’s also far less exciting than the headlines suggest.
What This Means
GPT-Rosalind matters less as a scientific breakthrough and more as a signal of where frontier AI is heading. The era of general-purpose models sold to everyone is giving way to domain-specific models sold to vetted buyers behind access controls. Future frontier models may increasingly arrive as locked rooms with badge readers rather than public products.
For drug discovery specifically, the question isn’t whether AI will accelerate research — it already has, in narrow ways. The question is whether that acceleration will benefit the research institutions that need it most, or primarily serve the pharmaceutical companies that can already afford decade-long development timelines.
OpenAI named this model after a scientist who never received proper credit for her contributions to one of history’s greatest discoveries. It would be ironic if the model bearing her name ended up primarily serving organizations that already have every advantage.