OpenAI shipped its first vertical AI model on April 16, 2026, and the target was not coders, customer service teams, or marketing departments. It was the $2.4 trillion pharmaceutical research industry. The model is called GPT-Rosalind, named after British chemist Rosalind Franklin, and it is a frontier reasoning model built to accelerate drug discovery, genomics analysis, protein reasoning, and translational medicine. The launch lands at a moment when pharma R&D productivity has been falling for three decades and the cost to bring a single drug to market now exceeds $2.6 billion, according to Tufts Center for the Study of Drug Development.
The release is the first in what OpenAI calls a broader life sciences model series, and it is gated behind a trusted-access program that requires a qualification and safety review before researchers can touch it. Early collaborators include Amgen, Moderna, the Allen Institute, Thermo Fisher Scientific, and Novo Nordisk, putting GPT-Rosalind in front of some of the most prolific biology labs on Earth from day one. The model achieved a 0.751 Pass@1 score on BixBench, the toughest open bioinformatics benchmark to date, and outperforms OpenAI’s general-purpose GPT-5.4 on 6 of 11 LABBench2 tasks, with the largest gain coming in CloningQA, an end-to-end test of designing DNA and enzyme reagents for molecular cloning protocols.
This article unpacks what GPT-Rosalind actually does, how it performs against incumbents like Google DeepMind’s AlphaFold 3 and Isomorphic Labs, why pharma giants moved so fast to sign up, and what the launch signals for the broader market for vertical AI models. We will look at benchmark numbers, partner use cases, expert reactions, regulatory questions, and the competitive landscape for AI in biology – including the awkward fact that OpenAI is now competing for the same revenue stream as one of its largest investors, Microsoft, whose Azure AI for Health platform already serves Pfizer, Merck, and Bayer.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Inside the April 16 launch: what GPT-Rosalind actually does
OpenAI announced GPT-Rosalind in a blog post on April 16, 2026, with formal media coverage from Fierce Biotech, Axios, and Drug Patent Watch following the next day. The product is described as a frontier reasoning model optimized for scientific workflows – not a chatbot, not an image generator, and not a coding copilot. It reasons over molecules, proteins, genes, biological pathways, and disease-relevant biology, and it is positioned to help researchers compress the multi-month grind between target identification and an experiment that actually works.
The model is available through three channels: ChatGPT, OpenAI’s coding agent Codex, and the API. Access is restricted to qualified customers under OpenAI’s trusted-access program, which requires applicants to clear a qualification and safety review covering biosecurity, dual-use research of concern, and provenance of training data. OpenAI did not publish per-token pricing for the research preview, a notable departure from the company’s usual practice of stamping a dollar figure on day one. The decision suggests that revenue is not the priority for this release; signing up flagship pharma labs and harvesting their feedback is.
OpenAI paired the model with a freely accessible Life Sciences research plugin for Codex that connects to more than 50 scientific tools and data sources, including sequence databases, structural biology repositories, and lab automation platforms. The plugin is the company’s first serious attempt to wire its agentic stack into the real-world scientific tool chain. In practical terms it means a researcher can ask GPT-Rosalind to design a cloning experiment, pull sequence data from a repository, draft a wet-lab protocol, and hand the work off to a robotic liquid handler – all within a single workflow.
The benchmark numbers: BixBench, LABBench2, and a 6-of-11 win over GPT-5.4
The headline number from the launch is a 0.751 Pass@1 score on BixBench, a benchmark designed around bioinformatics and data analysis tasks by Edison Scientific. BixBench tests whether a model can independently solve realistic computational biology problems, from RNA-seq quality control to variant calling – work that today consumes the time of mid-level computational biologists at most pharma companies. A 0.751 score puts GPT-Rosalind ahead of every published competing model and well clear of the human baseline reported by the benchmark’s authors.
The second benchmark, LABBench2, measures performance across 11 task families that mirror real laboratory workflows. GPT-Rosalind outperforms GPT-5.4 on 6 of 11 task families, with the largest improvement on CloningQA, which evaluates a model’s ability to design DNA and enzyme reagents for molecular cloning protocols end-to-end. CloningQA is the closest thing the benchmark world has to a real bench-scientist exam, and a meaningful gain there is the kind of result that pulls heads up in pharma R&D leadership meetings.
It is worth noting what the launch does not reveal. OpenAI has not published a score on ChemBench, GPQA-Diamond’s biology subset, MedQA, or any protein-structure benchmark like CASP. The company also did not disclose the model’s parameter count, training data composition, or the size of the biological corpus used for post-training. Drug Patent Watch flagged that GPT-Rosalind did not top GPT-5.4 on the other 5 LABBench2 task families, and noted there is still a significant gap for regulated R&D workflows where reproducibility and provenance are non-negotiable.
GPT-Rosalind vs the field: AlphaFold 3, Isomorphic Labs, and the new vertical race
GPT-Rosalind enters a market that already has serious incumbents. Google DeepMind’s AlphaFold 3, published in Nature in May 2024, predicts the structure and interactions of proteins, DNA, RNA, ligands, and other biomolecules with state-of-the-art accuracy. DeepMind’s drug-discovery spinout, Isomorphic Labs, has standing partnerships with Eli Lilly and Novartis worth a combined $3 billion in milestone payments. Microsoft’s Azure AI for Health serves Pfizer, Merck, and Bayer, and Anthropic has been quietly piloting a biology-tuned variant of Claude with at least three top-20 pharma companies.
The competitive read is that GPT-Rosalind is not trying to be AlphaFold 3. It is trying to be the reasoning layer that wraps around tools like AlphaFold 3, the lab notebooks, and the sequencing databases. Where AlphaFold 3 answers a narrow scientific question – given this sequence, what is the most likely structure? – GPT-Rosalind is built to plan, decompose, and orchestrate a research workflow that may include calling AlphaFold-style models as tools. That is a different and, in commercial terms, far larger market.
| Model | Provider | Type | Public benchmark | Access | Named pharma partners |
|---|---|---|---|---|---|
| GPT-Rosalind | OpenAI | Reasoning model + tool use | 0.751 BixBench Pass@1 | Gated research preview | Amgen, Moderna, Allen Institute, Thermo Fisher, Novo Nordisk |
| AlphaFold 3 | Google DeepMind | Structure prediction | Sub-Å RMSD on protein-ligand | AlphaFold Server (academic free) | Lilly, Novartis (via Isomorphic Labs) |
| Isomorphic Drug Discovery | Isomorphic Labs | End-to-end pipeline | Not publicly disclosed | Bespoke pharma deals | Eli Lilly ($1.7B), Novartis ($1.2B) |
| Claude biology variant | Anthropic | Reasoning model (pilot) | Not publicly disclosed | Private pilot | Three undisclosed top-20 pharma |
| Microsoft Azure AI for Health | Microsoft | Platform stack | Workload-specific | Commercial GA | Pfizer, Merck, Bayer |
| GPT-5.4 (general) | OpenAI | General reasoning | 5/11 LABBench2 wins vs Rosalind | Tier-based API | Used across pharma but not biology-tuned |
Inside the pharma partner list: who is doing what with GPT-Rosalind
OpenAI named five flagship collaborators in the launch announcement. Amgen got the most prominent quote, with Chief Information & Digital Officer Sean Bruich saying: “Our unique collaboration with OpenAI enables us to apply its most advanced capabilities and tools in new and innovative ways, with the potential to accelerate how we deliver medicines to patients.” Amgen has been working with OpenAI since 2024 and is using GPT-Rosalind across multiple research workflows from target identification through translational medicine. Amgen’s press office has published the partnership terms.
Moderna is the second high-profile name, building on an existing relationship that extends back to its early ChatGPT Enterprise rollout. The mRNA company is one of the heaviest internal users of OpenAI products in pharma, with reports that more than half its 6,000 employees use ChatGPT daily. Novo Nordisk, the Danish GLP-1 giant currently valued at over $400 billion, joined the partner list against the backdrop of soaring R&D budgets for next-generation obesity and metabolic disease drugs. Novo’s specific use cases were not disclosed but multiple sources point to obesity target deconvolution and on-target/off-target effect prediction.
Thermo Fisher Scientific brings a different angle: it is not a drug developer, it is the picks-and-shovels supplier to the entire industry, with annual revenue above $42 billion. Embedding GPT-Rosalind into Thermo Fisher’s instrument software and informatics products would put OpenAI’s biology model into thousands of labs that do not have a direct relationship with OpenAI. The Allen Institute rounds out the list and is the lone non-profit; it brings neuroscience, cell biology, and immunology datasets that are likely to feed back into future model versions.
The economics: why pharma is paying attention to an unpriced research preview
The pharmaceutical industry has a productivity problem so well documented it has its own name: Eroom’s Law, the inverse of Moore’s Law. The number of new drugs approved per billion dollars of R&D spending has been falling by roughly half every nine years since 1950. Today, average R&D spend per approved drug is around $2.6 billion, with development timelines stretching past 10 years. A 10% productivity gain across the industry’s roughly $280 billion in annual R&D outlay would unlock $28 billion of net value – and that is the prize sitting in plain sight for any model that can credibly accelerate target validation, candidate selection, or trial design.
OpenAI’s own commercial story is also leaning more heavily on vertical wins. The company recently closed a $122 billion raise at an $852 billion valuation and signed a $38 billion deal with AWS that ended its Azure-exclusivity arrangement. Both deals were predicated on revenue ramps that look increasingly difficult to hit through consumer ChatGPT growth alone. Vertical models in life sciences, legal, financial services, and manufacturing are the credible path to the kind of contracts that justify the company’s compute bill – and pharma has both the budgets and the willingness to pay top-tier per-seat enterprise pricing once value is proven.
The unspoken financial subplot is what happens to OpenAI’s relationship with Microsoft. Azure AI for Health is one of Microsoft’s most prominent vertical AI offerings and serves direct competitors to GPT-Rosalind’s launch partners. With OpenAI now selling to Amgen and Moderna directly, the two firms are formally competing for the same biology-AI dollars they once split – a structural tension that will play out in 2026 and 2027 enterprise procurement cycles.
Expert reactions: what scientists and analysts are saying
The reaction inside the scientific community has been more measured than the reception of consumer-facing OpenAI launches. Joy Jiao, who leads OpenAI’s life sciences research team, has been the main public voice for the release. In comments to Axios, Jiao emphasized that GPT-Rosalind is “not intended to replace scientists” but to assist with labor-intensive analytical tasks. Her framing – assistant, not replacement – is a deliberate echo of the language used by Google Health and Microsoft Research, and a recognition that the audience here is far more skeptical than a developer crowd.
Sean Bruich, Chief Information & Digital Officer at Amgen, said in the OpenAI announcement: “Our unique collaboration with OpenAI enables us to apply its most advanced capabilities and tools in new and innovative ways, with the potential to accelerate how we deliver medicines to patients.” Bruich’s quote is the only first-party pharma executive testimonial in the launch materials. Derek Lowe, the long-running In the Pipeline columnist and one of the most influential critical voices in pharma R&D, has cautioned that scientific benchmark scores must translate to validated wet-lab outcomes before any of this matters: “We have seen impressive AI numbers for years. The question is whether a single project gets to clinical readout faster because of this model.”
Derya Unutmaz, MD, professor at The Jackson Laboratory, has been an early public booster of GPT-Rosalind on X, posting workflow examples that he says compressed a multi-day literature review and experimental design exercise into roughly 30 minutes. Eric Topol, founder and director of the Scripps Research Translational Institute, was more guarded, telling Fierce Biotech-adjacent outlets that the model’s “real test is whether it can help avoid the false-positive enthusiasm that has dogged AI drug discovery since 2017.” Topol’s framing matters because his Scripps team has independently evaluated more biomedical AI models than almost anyone else in the field.
Regulatory, safety, and biosecurity questions
A biology-tuned frontier model raises a regulatory question that text generation models simply do not. Dual-use research of concern – DURC – covers experiments and capabilities that could be repurposed for harm, including pathogen enhancement and the synthesis of biological agents. OpenAI has stated that GPT-Rosalind is gated behind a trusted-access program with safety review precisely because of this risk. The company is using a layered approach: hard refusals on known biosecurity red lines, classifier-based monitoring on user inputs, and customer-level attestations covering institutional biosafety oversight.
Independent biosecurity researchers, including those at the Johns Hopkins Center for Health Security and the Nuclear Threat Initiative, have argued for years that any model capable of producing meaningful uplift in biological design must be evaluated against pre-published red-team criteria. So far OpenAI has not published the results of GPT-Rosalind’s biosecurity red-team evaluations, only stating that the trusted-access program is informed by them. The company’s recent willingness to push public-facing policy positions on AI’s broader societal impact suggests Rosalind-specific transparency could follow.
The FDA does not regulate AI tools that influence early-stage drug discovery, but it does regulate Software as a Medical Device and any model used in clinical decision-making. GPT-Rosalind is not currently positioned as a SaMD, but the moment a partner uses it to influence trial design or patient selection – both of which the launch material implicitly hints at – that calculus changes. In the EU, the AI Act’s high-risk classification for medical and pharmaceutical use is also expected to apply to any production deployment.
The historical context: from Rosalind Franklin to GPT-Rosalind
The naming choice is unusually deliberate. Rosalind Franklin was the British chemist whose X-ray diffraction images – most famously Photo 51 – provided the critical experimental evidence for the double-helix structure of DNA. Her contributions were under-credited during her lifetime, with Watson, Crick, and Wilkins sharing the 1962 Nobel Prize for the structure she helped reveal. Naming a frontier biology model after her is part homage and part marketing: an explicit signal that OpenAI wants this product to land in academic and translational research culture, not just industrial pharma.
The historical parallel is not coincidental. Photo 51 was a single experimental input that unlocked decades of follow-on biology. GPT-Rosalind’s pitch is that a sufficiently capable reasoning layer over biological data could play a similar catalytic role today – not by replacing the experimentalists, but by compressing the cycle time between observation, hypothesis, and the next experiment. Whether that pitch survives contact with the messy realities of wet-lab science is precisely what the next 12 months of partner pilots will reveal.
Stock and market impact: pharma, AI infrastructure, and the vertical model thesis
The April 16 launch did not move pharma indexes in any dramatic way – markets have learned to discount AI announcements until pipeline data confirms them. Amgen and Moderna both saw modest single-digit intraday moves on positive sentiment, while Nvidia closed up roughly 1.4% on the day on broader AI demand. The more meaningful signal was in the venture market: AI-for-biology startups including Recursion Pharmaceuticals, Insitro, and Cradle Bio all saw renewed inbound from generalist VCs the week after the launch, on the read that OpenAI’s entry validates the category rather than crushes it.
The macro thesis is straightforward. If GPT-Rosalind works, it accelerates the shift toward vertical AI models that OpenAI’s competitors – Anthropic, Google, and a long tail of specialists – will all have to match. If it does not, it becomes the highest-profile data point in the case that general-purpose frontier models cannot easily be retargeted at high-stakes scientific domains. Either outcome is informative, which is part of why the Anthropic vs OpenAI rivalry has become the most-watched commercial contest in AI.
| Metric | Pharma R&D 2026 | What GPT-Rosalind targets |
|---|---|---|
| Average R&D cost per approved drug | ~$2.6B (Tufts CSDD) | Cycle-time compression in target ID and candidate selection |
| Average development time | ~10–12 years | Faster experimental planning + protocol design |
| Phase 1 to approval success rate | ~7.9% (BIO/Informa) | Better target validation, fewer dead-end programs |
| Annual global pharma R&D spend | ~$280B (Evaluate Pharma) | Productivity gains across the entire spend base |
| Computational biologists per top-20 pharma | 200–600 typical | Force multiplier, not replacement |
| Wet-lab cycle time per experiment | 2–6 weeks typical | Reagent design (CloningQA gain) |
| AI for drug discovery market size | ~$3.5B (2025) → $13B (2030) | Direct addressable market |
How GPT-Rosalind compares to OpenAI’s other 2026 launches
GPT-Rosalind is the third major OpenAI product release of 2026, following the rollout of the GPT-5.4 family and the new agentic Codex capabilities. It is the first time the company has launched a model that is not a general-purpose chatbot or a developer tool, and the first to ship inside a trusted-access program from day one. The choice is a strategic departure from OpenAI’s historical playbook, which has been to ship broadly, monetize through tiers, and iterate. With Rosalind, the company is saying that the value of getting the science right with five trusted partners outweighs the optics of a broad launch.
The comparison to Claude vs ChatGPT in the general AI race is instructive. In general reasoning, OpenAI and Anthropic are within a few percentage points of each other on SWE-bench and other benchmarks. 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, and why Amgen, Moderna, and Novo Nordisk all wanted to be named on day one rather than wait for the next release cycle.
The vertical AI model thesis: what GPT-Rosalind tells us about 2026
The pattern is now becoming clearer. DeepSeek’s MATH-500 dominance showed that targeted post-training can push specific capabilities well ahead of general-purpose frontier models. OpenAI’s revenue miss earlier this year made clear that the consumer ChatGPT business cannot single-handedly support the company’s compute roadmap. The vertical AI playbook – take a frontier model, post-train it on a high-value domain, gate it behind a trusted-access program, sign flagship partners, and grow into the enterprise contract – is increasingly looking like the path that justifies the capex.
The expected next moves: a legal-services model, a financial-services model, and a manufacturing model from OpenAI within the next 12 months. Anthropic has telegraphed similar plans, and Google DeepMind continues to publish a steady stream of domain-specific papers from its Health team. The risk is fragmentation – pharma CIOs do not want to manage 17 different gated AI relationships – but the early signal is that the leading labs would rather own the high-value domain stacks than cede them to vertical-only startups.
Five predictions for GPT-Rosalind and AI in life sciences
1. A publicly disclosed validation study by end of 2026. Expect Amgen or Moderna to publish a preprint or peer-reviewed paper using GPT-Rosalind as a co-author or a methods component within nine months. The competitive incentive to show real wet-lab impact is too strong, and the first paper sets the narrative.
2. Anthropic will counter with a named biology model by Q3 2026. The company’s pilot Claude biology variant is already in front of pharma partners, and the launch playbook is now clear. Watch for a named release – possibly “Claude Crick” or similar – paired with at least one top-10 pharma partner.
3. GPT-Rosalind goes paid-tier within 12 months. The unpriced research preview is a customer-acquisition exercise, not a long-term plan. Expect a tiered enterprise API offering with per-seat pricing in the $1,000–$5,000/month range, plus a usage-based component for high-volume API calls.
4. At least one biosecurity incident – successfully blocked – gets disclosed. Frontier labs have started publishing case studies of red-teamed attempts to misuse biology models. With trusted access in place, OpenAI is well positioned to share a controlled example, both to demonstrate safety controls and to set the public benchmark for the industry.
5. The FDA will issue draft guidance on AI in drug discovery within 18 months. The agency has been preparing for vertical AI in pharma since 2023 and the GPT-Rosalind launch is the kind of forcing function that accelerates a public draft. Expect requirements around model provenance, validation evidence, and bias testing for any model used in trial design or biomarker discovery.
How researchers can apply for access today
GPT-Rosalind access is gated through OpenAI’s trusted-access program. Eligible applicants must be affiliated with a recognized research institution or pharmaceutical/biotech company, agree to OpenAI’s usage policies including biosecurity restrictions, and pass a qualification review covering institutional biosafety oversight. The free Life Sciences research plugin for Codex is available without going through trusted access and is a reasonable on-ramp for individual researchers who want to evaluate the toolchain before committing.
# Example API call (pseudocode, requires trusted-access credentials)
from openai import OpenAI
client = OpenAI(api_key="sk-rosalind-trusted-access-key")
response = client.chat.completions.create(
model="gpt-rosalind-preview",
messages=[
{"role": "system", "content": "You are an expert molecular biologist."},
{"role": "user", "content": "Design a Gibson assembly cloning protocol "
"to insert the human BRCA1 cDNA into pcDNA3.1 "
"with a C-terminal FLAG tag. Provide primer "
"sequences, restriction enzyme alternatives, "
"and a step-by-step lab protocol."}
],
tools=[{"type": "life_sciences_research"}]
)
print(response.choices[0].message.content)For institutional users, OpenAI is routing applications through the Enterprise team and is prioritizing applicants with existing institutional biosafety committees and clear use cases that fall outside the published DURC restrictions. Public universities, NIH-funded centers, and named pharma R&D groups are the first cohort being onboarded. Additional technical detail on benchmarks, partner workflows, and the trusted-access criteria has been collected by Drug Patent Watch and PharmaNow.
Frequently asked questions
What is GPT-Rosalind?
GPT-Rosalind is OpenAI’s frontier reasoning model for life sciences research, announced on April 16, 2026. It is built to accelerate drug discovery, genomics analysis, protein reasoning, and translational medicine, and is available as a gated research preview through ChatGPT, Codex, and the OpenAI API.
How does GPT-Rosalind compare to AlphaFold 3?
They solve different problems. AlphaFold 3 predicts the 3D structure of biomolecules and their interactions. 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 for tasks like literature synthesis, experimental design, and reagent generation.
Who are the named GPT-Rosalind launch partners?
OpenAI named Amgen, Moderna, the Allen Institute, Thermo Fisher Scientific, and Novo Nordisk as early collaborators on the GPT-Rosalind launch. Eli Lilly has a broader prior relationship with OpenAI that predates the Rosalind release.
What benchmark scores has OpenAI published?
The headline number is a 0.751 Pass@1 score on BixBench, the bioinformatics and data-analysis benchmark from Edison Scientific. On LABBench2, GPT-Rosalind outperforms GPT-5.4 on 6 of 11 task families, with the largest gain on CloningQA, an end-to-end molecular cloning reagent design task.
Is GPT-Rosalind paid?
OpenAI did not publish pricing for the April 16 research preview. Access is gated to qualified customers under the trusted-access program. The accompanying Life Sciences research plugin for Codex is freely accessible and connects to over 50 scientific tools and data sources.
Why is GPT-Rosalind named after Rosalind Franklin?
The model is named for the British chemist whose X-ray diffraction work, including Photo 51, provided the experimental evidence behind the double-helix structure of DNA. The naming positions GPT-Rosalind as a tool intended to serve foundational biology research, not just industrial drug discovery.
What are the safety controls on GPT-Rosalind?
OpenAI is using a layered approach: trusted-access gating with a qualification and safety review, hard refusals on known biosecurity red lines, classifier-based input monitoring, and customer-level attestations covering institutional biosafety oversight. The full red-team evaluation results have not yet been published.
Will GPT-Rosalind replace computational biologists?
OpenAI explicitly positions it as a force multiplier rather than a replacement. Joy Jiao, who leads OpenAI’s life sciences research, has emphasized that the model is intended to assist scientists with labor-intensive analytical tasks, not to substitute for them. Real-world impact will depend on validated wet-lab outcomes over the coming year.
What is BixBench?
BixBench is a benchmark developed by Edison Scientific that evaluates AI models on realistic bioinformatics and computational biology tasks, including RNA-seq quality control, variant calling, and sequence analysis. GPT-Rosalind’s 0.751 Pass@1 score is the highest publicly reported result on the benchmark to date.
How does GPT-Rosalind affect Microsoft’s Azure AI for Health?
OpenAI is now selling directly to flagship pharma customers including Amgen and Moderna, putting the company in competition with Microsoft’s Azure AI for Health platform, which serves Pfizer, Merck, and Bayer. The dynamic is part of a wider structural shift since OpenAI’s $38 billion AWS deal ended Azure-exclusive cloud hosting earlier this year.
What to watch in the next 90 days
The GPT-Rosalind launch is the start of a longer 2026 storyline rather than a self-contained event. Three near-term inflection points are worth tracking. First, OpenAI’s safety disclosures: the company committed to publishing more detailed red-team results, and the specifics will shape regulator expectations across the industry. Second, the partner ecosystem: a single high-credibility preprint with a named experimental outcome from Amgen, Moderna, or the Allen Institute will move the conversation from benchmark scores to validated impact. Third, the competitive response from Anthropic, Google DeepMind, and the long tail of biology-specific startups.
For pharma decision-makers, the practical question is whether to begin formal procurement evaluations now or wait for the priced enterprise tier. The smart play for top-20 pharma is almost certainly to onboard via trusted access now – the cost of standing pilots is low compared with being last to optimize internal workflows. For investors, the signal to watch is how quickly the Rosalind blueprint propagates to other verticals, and how the legal-services, financial-services, and manufacturing equivalents land within the next four quarters. For working scientists, the only honest answer is that the next 12 months of wet-lab data will determine whether GPT-Rosalind is the start of a new era or another high-profile benchmark trophy.
