Rosalind Franklin took Photograph 51 in May 1952. It was an X-ray diffraction image of hydrated DNA. The image is the reason we know the structure of the double helix, even though James Watson and Francis Crick are the names most people associate with the discovery. Seventy-four years later, OpenAI has named a model after her.
GPT-Rosalind, announced on Thursday, April 16, is OpenAI's first reasoning model trained specifically for life sciences research. It is not a general chatbot with a biology hat on. It is a frontier model tuned for evidence synthesis, hypothesis generation, experimental planning, and multi-step reasoning across biochemistry, genomics, and translational medicine. It can query specialized biological databases, parse scientific literature, run computational tools, and propose experimental pathways inside a single workflow.
The list of companies already using it reads like a who's who of modern biotech. Amgen. Moderna. Novo Nordisk. Thermo Fisher Scientific. Oracle Health and Life Sciences. NVIDIA. The Allen Institute. Benchling. The UCSF School of Pharmacy. OpenAI says qualified enterprise customers in the United States can apply to access the model through a program it calls Trusted Access.
Nobody outside that program will be able to use it.
The Model Is Designed Around One Job
Most of OpenAI's recent launches have been horizontal. GPT-5.4 is a better generalist. Codex is a better coding agent. GPT-Rosalind is the company's first frontier model tuned for a single vertical, and the choice of which vertical is telling. The money to fund the specialty bet is not in question: OpenAI closed a $122 billion funding round on March 31, with Amazon writing the single largest check at 50 billion dollars.
The life sciences research workflow is exactly the kind of problem where a general-purpose model hits a wall. It requires reading hundreds of papers, cross-referencing genomic databases, reasoning about protein structure, designing experiments that cost millions of dollars to run, and explaining the reasoning to a reviewer who will catch any hand-wave. OpenAI's life sciences research lead, Joy Jiao, told Bloomberg the model was designed to be "better at fundamental reasoning" in fields like biochemistry and genomics, where the failure mode of a wrong answer is not embarrassment but a wasted six-month trial.
Its benchmark scores point in the same direction. On BixBench, a public test built around real-world bioinformatics and data analysis tasks, GPT-Rosalind posted a pass rate of 0.751. The leaderboard as reported by OpenAI at launch:
| Model | BixBench Pass Rate |
|---|---|
| GPT-Rosalind | 0.751 |
| GPT-5.4 | 0.732 |
| GPT-5 | 0.728 |
| Grok 4.2 | 0.698 |
| Gemini 3.1 Pro | 0.550 |
A roughly 2 percentage point gain over GPT-5.4 sounds small. In a benchmark where each percentage point represents the difference between a wasted experimental design and a viable one, it is not small. It is the reason Thermo Fisher and Novo Nordisk signed up early.
The Partners Are the Product Story
OpenAI published executive quotes from two of the early customers on Thursday.
Sean Bruich, SVP of AI and Data at Amgen, said the collaboration lets the company apply advanced reasoning tools in ways that could "accelerate how we deliver medicines to patients." Amgen has historically been a heavy spender on computational biology, and the relationship with OpenAI puts the model inside workflows that previously ran on internal tooling.
Stéphane Bancel, CEO of Moderna, framed the model's strength as the ability to "reason across complex biological evidence" and translate insight into experimental workflows. Moderna has deployed generative AI inside its mRNA pipeline since 2023, and the partnership places GPT-Rosalind alongside those internal systems rather than replacing them.
The other named customers cover the full research stack. Novo Nordisk is a leader in peptide and protein drug development. Thermo Fisher sells the instruments and reagents that actually run the experiments. Oracle Health and Life Sciences is the infrastructure layer underneath many large pharma companies' research data. NVIDIA is both a partner and a symbolic choice, given its prior deep investment in AI for drug discovery through BioNeMo. The Allen Institute runs some of the most ambitious public neuroscience and biology datasets in the world. Benchling is the electronic lab notebook of choice for modern biotech startups. UCSF School of Pharmacy brings an academic research frontier.
The partner list serves a second purpose beyond distribution. It is the qualification bar. Organizations applying to use GPT-Rosalind have to demonstrate legitimate scientific intent and the governance infrastructure to prevent misuse. A new biotech running out of a garage cannot get in. A research university with an institutional biosafety committee can.
Access Is Deliberately Narrow for a Specific Reason
OpenAI built GPT-Rosalind on three stated principles: beneficial use, strong governance, and controlled access. The product of those principles is a Trusted Access program that explicitly excludes the general public, excludes most enterprise customers, and requires a qualification review for every organization that wants in.
That restriction is a direct response to a threat model that has been published in peer-reviewed journals. A Science paper titled "Strengthening nucleic acid biosecurity screening against generative protein design tools," published in the final week of 2025, documented a specific vulnerability. AI protein design systems can produce variants of dangerous proteins whose DNA sequences evade the biosecurity filters that nucleic acid synthesis providers use to screen commercial orders. The paper's authors worked with screening vendors to deploy patches, and detection rates improved. But the mere fact that the paper had to be written is the reason GPT-Rosalind is not a public API.
A model that can reason about "how to produce a protein variant that retains wild-type-like function" is the same kind of model that, without governance, can reason about how to produce a dangerous protein variant that evades detection. OpenAI is not pretending the risk is zero. It is pretending the risk is manageable if the access list is managed.
The Competitive Framing Is Anthropic and Google DeepMind
OpenAI is the third frontier lab to plant a flag specifically in life sciences.
Anthropic's response has been a different shape. The company's unreleased Mythos frontier model, previewed in late March, found thousands of zero-day vulnerabilities during Project Glasswing, and Anthropic has pitched it to defense and security customers before biotech. The cyber market is where Anthropic is first.
Google's answer in life sciences is older and deeper. AlphaFold, trained by DeepMind, solved the protein structure prediction problem at a scale that won its creators a share of the 2024 Nobel Prize in Chemistry. Isomorphic Labs, the DeepMind spinout, has signed drug discovery deals with Eli Lilly and Novartis. Google's approach has been to build foundational scientific models and spin them into dedicated companies.
OpenAI's approach is the third path. Instead of building a single scientific model or spinning out a research company, it is shipping a frontier reasoning model tuned for science and selling it, through controlled access, directly into the research workflows of the biggest customers in the category.
The bet is that the best specialized reasoning model wins, even if Anthropic has the more dramatic cybersecurity story and Google has the Nobel-grade protein model. Drug discovery is not about a single breakthrough. It is about a thousand decisions made faster and with fewer errors across a ten-year development cycle. A model that does each of those decisions a little better is worth billions in shortened timelines.
The Other Side
Several critics pushed back immediately on Thursday.
The first critique is performance-inflation. The BixBench leaderboard is curated by OpenAI, and the pass rates were reported by OpenAI at launch. Independent confirmation of the scoring will take weeks. Researchers who have worked with domain-specific benchmarks before noted that a 2-point gap is exactly the kind of delta that evaporates when a third party re-runs the evaluation with different seeds and prompts. The lesson from Humanity's Last Exam, the reasoning benchmark that has held up against every frontier model this year, is that published scores need independent replication before they are worth trusting.
The second critique is biosecurity theater. The Trusted Access gating looks rigorous, but qualified enterprise customers have large employee bases. Anyone with access inside Amgen or Moderna can, in practice, query the model. A leak of prompts, outputs, or the model weights themselves would erase the governance layer in a single day. Critics argued the correct control point is at the nucleic acid synthesis step and at downstream biological containment, not at the reasoning layer.
The third critique is commercial. A model that is only available to the largest biotech incumbents reinforces the existing power structure in drug development. Small and mid-size labs, which have historically driven a disproportionate share of breakthrough science, will not be able to access the same reasoning capability their larger competitors just gained. That is the opposite of what open foundational science has historically looked like.
Supporters inside the biotech partner list pushed back that this is the realistic path. A fully open biological reasoning model is a proliferation risk they are not willing to accept. A closed model with audited customers is, in their framing, the only version of this technology that can ship at all.
The Bottom Line
GPT-Rosalind is OpenAI's admission that horizontal scaling has limits and its first bet on vertical specialization. Biology was the right place to plant the flag. It is also, because of biosecurity risk, the last place OpenAI will ship a model with an open API. The strategy for every future specialty model is now visible: pick a vertical with real business value, train a frontier reasoning model against its benchmarks, sign the ten most important customers, and gate access by governance review.
For an ML engineer watching this release, the practical takeaway is quieter than the Thursday headlines. Specialty frontier models are now shipping at the same cadence as general-purpose models. The next five years of AI is not going to be one model that does everything. It is going to be a dozen models that each do one thing to a reasoning standard the general model cannot match, sold through controlled-access programs that look more like Pharma procurement than SaaS sign-up.
The question worth asking before the next OpenAI specialty launch is simpler than the benchmark scores. Which customer list tells you what the product actually is? On April 16, the list said Amgen, Moderna, Novo Nordisk, and Thermo Fisher. That is the product.
Sources
- OpenAI Takes on Google With New AI Model Aimed at Drug Discovery — Bloomberg (April 16, 2026)
- OpenAI launches new AI model for life sciences research — Axios (April 16, 2026)
- OpenAI debuts GPT-Rosalind, a new limited access model for life sciences — VentureBeat (April 16, 2026)
- OpenAI Launches GPT-Rosalind: Its First Life Sciences AI Model Built to Accelerate Drug Discovery and Genomics Research — MarkTechPost (April 16, 2026)
- OpenAI introduces GPT-Rosalind, its drug discovery AI — pharmaphorum (April 17, 2026)
- OpenAI launches biopharma-focused AI model to compete with Anthropic — Endpoints News (April 16, 2026)
- OpenAI launches GPT-Rosalind, a reasoning model built for life sciences research — The Decoder (April 17, 2026)
- OpenAI Unveils GPT-Rosalind, Expands AI Push into Biotech and Pharma R&D — Medical Dialogues (April 17, 2026)
- OpenAI launches GPT-Rosalind, a specialised AI model for drug discovery and life sciences research — The Next Web (April 16, 2026)
- Strengthening nucleic acid biosecurity screening against generative protein design tools — Science (December 2025)
- OpenAI Launches AI Model GPT-Rosalind for Life Sciences Research — US News (April 16, 2026)