GPT-Rosalind: OpenAI's 2026 Life Sciences AI Model

On April 16, 2026, OpenAI introduced GPT-Rosalind, its first frontier reasoning model built exclusively for biology, drug discovery, and translational medicine.¹ The model is available only as a research preview to qualified US enterprise customers through a Trusted Access Program, and it arrives alongside a free Life Sciences research plugin for Codex that connects OpenAI's mainline models to more than 50 public biological databases.² Named after British chemist Rosalind Franklin, whose X-ray diffraction work helped reveal the double-helix structure of DNA, the model marks OpenAI's first purpose-built domain-specific release and a visible pivot away from one-model-fits-everything toward vertical AI.³

What You'll Learn

• What GPT-Rosalind is and how it differs from general-purpose GPT-5.4
• The benchmark numbers on BixBench and LABBench2 and how they compare with frontier rivals
• Who the launch partners are and how the Trusted Access Program works
• How the free Life Sciences plugin for Codex fits into the rollout
• How GPT-Rosalind stacks up against Google's Isomorphic Labs and Chai Discovery
• The biosecurity concerns and safeguards OpenAI built into the deployment

TL;DR

GPT-Rosalind is OpenAI's first vertical, science-tuned reasoning model. It hit a pass@1 score of 0.751 on BixBench, a comprehensive bioinformatics benchmark, ahead of GPT-5.4 (0.732), GPT-5 (0.728), Grok 4.2 (0.698) and Gemini 3.1 Pro (0.550).⁴ On LABBench2, a 2026 benchmark spanning nearly 1,900 biology research tasks, GPT-Rosalind beat GPT-5.4 on 6 of 11 task families, with the biggest gain in CloningQA.⁵ Access is restricted: research preview only, US enterprise customers only, through a trusted-access program. Launch partners include Amgen, Moderna, Thermo Fisher Scientific, the Allen Institute, Oracle Health and Life Sciences, NVIDIA, Benchling and UCSF School of Pharmacy.⁶

Why OpenAI Built a Biology-Specific Model

For three years, the dominant AI playbook has been scale up one model, let it do everything. GPT-Rosalind is a public admission that this approach has limits in science. The model was built to handle the work that chews up most of a biologist's week: synthesizing hundreds of recent papers, parsing sequencing outputs, planning CRISPR experiments, and chaining together tools that live in separate silos. In OpenAI's framing, GPT-Rosalind is positioned as a "research partner" — a phrase used by Joy Jiao, OpenAI's head of life sciences research, to describe how the model should plug into corporate R&D teams.⁷

The strategic context matters. OpenAI has recently shipped a string of vertical plays: a Hiro Finance acquisition for financial reasoning, a broader Codex enterprise push, and now a science-only frontier model. GPT-Rosalind is the first in what OpenAI calls a new family of domain-optimized models, and it positions the company directly against Google DeepMind's Isomorphic Labs spinoff, which has already secured roughly $3 billion in deals with Eli Lilly and Novartis to co-design drug candidates using AlphaFold 3 and its successor engines.⁸

Benchmark Performance on BixBench and LABBench2

Two benchmarks anchor GPT-Rosalind's scientific credentials: BixBench and LABBench2.

BixBench, introduced by researchers from FutureHouse and ScienceMachine in early 2025 and now maintained by Edison Scientific, is a comprehensive test of bioinformatics agents. It presents 53 real-world analytical scenarios and 296 associated questions, handing an agent an empty Jupyter notebook, raw data files and freedom to plan its own analysis.⁹ On BixBench, GPT-Rosalind hit 0.751 pass@1, ahead of every tested frontier model.

Source: reported in OpenAI's announcement coverage.⁴

LABBench2 is a 2026 upgrade to the original LAB-Bench, co-developed by Edison Scientific and building on FutureHouse's earlier dataset. It comprises nearly 1,900 tasks covering literature understanding and retrieval, database access, sequence manipulation, protocol troubleshooting, molecular biology assistance, and experiment planning. On LABBench2, GPT-Rosalind outperformed GPT-5.4 on 6 of 11 task families, with its largest margin on CloningQA, which requires end-to-end design of DNA and enzyme reagents for molecular cloning protocols.⁵

The benchmarks tell a consistent story. GPT-Rosalind is not a general-purpose upgrade over GPT-5.4; on broad reasoning or coding, it is not claimed to lead. What it is better at is the bioinformatics and wet-lab planning work where GPT-5.4 already had strong capability and Rosalind adds a few percentage points of headroom — the kind of margin that, multiplied across a pharma pipeline, is the difference between a candidate that moves into IND-enabling studies and one that dies in screening.

The Life Sciences Plugin for Codex

Alongside GPT-Rosalind, OpenAI quietly shipped what may be the more immediately useful piece for rank-and-file researchers: a free Life Sciences research plugin for Codex. Unlike the gated model, the plugin is unrestricted and works with OpenAI's mainline models.¹⁰

The plugin bundles modular skills for research workflows and connects models to more than 50 public multi-omics databases and literature sources, including AlphaFold for protein structure lookup, Bgee for gene expression data, and BindingDB for ligand–target affinity. It is designed to handle common research chores: protein structure lookup, sequence search, literature review, and public dataset discovery. A solo computational biologist working on a tight grant budget can pick up most of GPT-Rosalind's workflow advantages for the price of a Codex subscription — a point the OpenAI team appears to be making deliberately.

Launch Partners and Trusted Access

GPT-Rosalind launched with a roster built from the full stack of biopharma and scientific infrastructure:

• Amgen (NASDAQ: AMGN) — biologics and small molecules
• Moderna (NASDAQ: MRNA) — mRNA vaccines and therapeutics
• Thermo Fisher Scientific — instruments and lab supplies
• Allen Institute — open neuroscience and bioscience research
• Oracle Health and Life Sciences — clinical data and EHR infrastructure
• NVIDIA — compute and GPU-accelerated biology tooling
• Benchling — R&D cloud platform used across biotech
• UCSF School of Pharmacy — academic research anchor

Source: multiple coverage reports of the launch.⁶

Access is narrow by design. To qualify for the Trusted Access Program, an organization must satisfy three requirements according to OpenAI's terms: the work has to be legitimate scientific research with clear public benefit; the organization needs proper governance, compliance, and abuse-prevention controls in place; and access has to be limited to approved users operating in secure, managed environments.¹¹ During the preview phase, the model does not consume existing API credits or tokens, a deliberate choice to make experimentation frictionless for scientists while OpenAI finalizes pricing.

How GPT-Rosalind Compares to Isomorphic Labs and Chai Discovery

OpenAI is not the first big lab to chase drug discovery. The competitive map in April 2026 is crowded.

Sources: Isomorphic Labs partnership coverage,⁸ Chai-2 performance announcement,¹³ Claude for Life Sciences launch,¹² Claude Opus 4.7 release notes.¹⁴

The distinction worth understanding is that Isomorphic Labs sells outcomes; OpenAI sells access. Isomorphic is a pharma co-development partner that uses its own stack to hand back candidates under commercial terms. GPT-Rosalind is, in contrast, a tool that enterprise biologists can wire into their own pipelines — provided they clear the Trusted Access bar. Chai Discovery occupies yet a third position: it ships antibody-specific generative models and runs its own wet-lab validation. A serious drug-discovery team in 2026 will likely use more than one of these — the stacks are more complementary than competitive at the workflow level.

Safety, Dual Use, and Biosecurity

Any frontier model capable of designing reagents or interpreting genomic data carries dual-use risk. OpenAI's response with GPT-Rosalind has three components. First, the model is gated — no consumer access, no general API availability, only approved enterprise users. Second, it is tied to technical safeguards that flag potentially dangerous activity and limits on how the model can be used. Third, every access request flows through an application-based review designed to screen out requests that do not satisfy the Trusted Access criteria.¹⁵

Critics note that gating is a deployment control, not a capability control — the underlying model exists, and the real biosecurity test will come when comparable capability lands in open-weight systems. Anthropic's 2025 red-team research on Claude 4 found that participants assisted by frontier models produced substantially better bioweapons-acquisition plans than the internet-only control group, and the company moved Opus 4 to AI Safety Level 3 in response.¹⁶ That result cuts in both directions for Rosalind: it validates OpenAI's decision to gate the model, and it raises the stakes as science-tuned capability continues to diffuse across labs.

What This Means for Biotech and Pharma Teams

For a mid-sized biotech that qualifies for trusted access, GPT-Rosalind offers three practical gains. It shortens literature review from days to hours for a focused target class. It turns protocol planning into a conversational loop with a model that actually knows what restriction enzymes do. And it plugs into Codex workflows, which means a bioinformatics engineer can ask the same model to write an analysis pipeline, run it against a public dataset, and draft the methods section of a paper all in one session.

For academic labs outside the Trusted Access bubble, the more immediately useful tool is the free Life Sciences plugin for Codex. It does not carry Rosalind's domain fine-tune, but it carries the 50-plus tool integrations, which is where most day-to-day friction lives.

The Executive Exit

A day after the launch, OpenAI announced that Kevin Weil, VP of OpenAI for Science and the executive who led GPT-Rosalind, was leaving the company along with Bill Peebles (the researcher behind Sora, whose app is being discontinued on April 26, 2026) and Srinivas Narayanan (CTO of B2B Applications).¹⁷ Weil's exit does not change GPT-Rosalind's trajectory — the model and the plugin shipped — but it does underscore how much internal churn sits behind OpenAI's enterprise pivot. Investors preparing for the IPO will read the shakeup both ways: as a signal that OpenAI is shedding consumer moonshots to focus on enterprise revenue, and as a reminder that the team that builds a science-AI practice is not necessarily the team that scales it.

Bottom Line

GPT-Rosalind is OpenAI's first real vertical model — a signal that the "one model rules them all" era is thinning out and that the next cycle of frontier AI will be fought in specialized domains. For qualified US biotech and pharma teams, the benchmarks are credible and the tool integration is genuinely useful. For the rest of the industry, the free Codex plugin is the more immediate win, and the bigger story is the direction of travel: Google DeepMind has Isomorphic Labs, OpenAI now has Rosalind, and the next six quarters will decide which framework pharma teams actually standardize on.

Sources