OpenAI is planting a very clear flag in biotech with GPT-Rosalind, a new version of its AI model built specifically for life sciences research rather than general chat or coding help. It is pitched as a way to speed up the slowest, most expensive part of drug development: turning messy biological data into testable hypotheses and better experiments.

If you work in biology or pharma, you already know the depressing stats. Getting a new drug from an early target to FDA approval can easily take 10 to 15 years in the United States, with billions of dollars burned along the way and a brutal failure rate in the clinic. OpenAI’s argument is simple: the early discovery pipeline is full of reading, cross-referencing databases, interpreting noisy experimental results, and designing the next round of experiments – exactly the kind of multi-step, information-heavy reasoning that frontier language models are getting very good at. GPT-Rosalind is OpenAI’s attempt to turn that general capability into a lab-ready assistant.

The model is named after Rosalind Franklin, the British scientist whose X-ray diffraction work was critical to uncovering the double-helix structure of DNA, and the branding is not subtle. OpenAI is aligning itself with the idea of rigorous, evidence-driven molecular biology – not just generic AI hype – and making a bet that the next wave of breakthroughs in health will come from scientists working hand in hand with domain-tuned models.

Under the hood, GPT-Rosalind is part of a dedicated “life sciences model series” optimized for tasks like reasoning over molecules, proteins, genes, pathways, and disease biology, rather than writing emails or summarizing a Netflix script. In OpenAI’s internal evaluations, it beats the company’s own GPT-5 family on chemistry understanding, protein analysis, experimental design, and tool usage – essentially the skills you need to navigate modern wet-lab and computational pipelines. It is also wired to use external software and databases more effectively, which is critical, because serious biology happens in tools like structure predictors, sequence search engines, and specialized omics databases, not just text.

On benchmarks, OpenAI came armed with numbers, because at this point, every AI launch lives or dies on a scoreboard. On BixBench, a bioinformatics and data-analysis benchmark built to look more like real lab work than trivia, GPT-Rosalind posts a pass rate around 0.75, ahead of other models with published scores, including its own GPT-5.4 and Google’s Gemini 3.1 Pro. On LABBench2, which measures tasks like literature retrieval, database access, sequence manipulation, and protocol design, OpenAI says Rosalind outperforms GPT-5.4 on 6 out of 11 tasks, with its biggest gain on CloningQA – a test that involves designing the DNA constructs and enzymes for molecular cloning end to end.

The more eye-catching stat comes from a collaboration with Dyno Therapeutics, a company that uses AI to design gene therapy capsids. OpenAI reports that on an RNA sequence-to-function prediction task using unpublished sequences, best-of-ten runs from GPT-Rosalind ranked above the 95th percentile of 57 human experts in the AI-bio field; on a sequence generation task, it landed around the 84th percentile. That does not mean Rosalind suddenly replaces scientists or guarantees better drugs, but it does suggest that, in very specific, well-defined tasks, the model can perform at or above the level of strong domain specialists.

The bigger story here is how OpenAI wants people to actually use this thing. GPT-Rosalind is meant less as a standalone chat interface and more as a reasoning layer that sits on top of scientific data sources and tools. Alongside the model, the company is shipping a Life Sciences research plugin for Codex – essentially a modular toolkit that connects the model to more than 50 multi-omics databases, literature repositories, and biology utilities. Out of the box, this plugin supports workflows like protein structure lookup, sequence similarity searches, literature review, and public dataset discovery, so a typical lab question might chain through multiple tools automatically: find variants in a gene, pull structural context, cross-reference disease associations, and assemble a reading list of key papers in one shot.

OpenAI describes these plugin “skills” as an orchestration layer for broad, ambiguous research questions, where the model has to decompose the task, call the right databases, and then synthesize results into something a scientist can act on. This is the place where domain-specific AI can move beyond pretty summaries. In principle, a biologist could use Rosalind to troubleshoot a stubborn reaction, design a set of CRISPR gRNAs, or reanalyze a confusing omics dataset, with the model not just chatting, but actually parsing experimental tables, spotting patterns, and proposing follow-up experiments grounded in both data and literature.

Access, however, is deliberately not wide open. GPT-Rosalind is launching as a research preview for “qualified” customers through a trusted-access program, starting with Enterprise users in the United States. Organizations have to request access, pass a qualification and safety review, and agree to additional life sciences research terms on top of OpenAI’s standard usage policies. OpenAI says it is evaluating applicants on three core principles: beneficial use, strong governance and safety oversight, and tightly controlled access with enterprise-grade security. In practice, that means they are looking for institutions doing legitimate research with clear public benefit, capable of maintaining biosafety and misuse-prevention controls, and able to restrict model use to vetted users in secure environments.

To underscore its seriousness – and to gain real-world feedback – OpenAI has lined up a set of flagship partners, including pharma heavyweights like Amgen and Moderna, life sciences suppliers like Thermo Fisher Scientific, and research groups such as the Allen Institute and UCSF’s School of Pharmacy. Amgen’s AI and data chief Sean Bruich is quoted pointing to the “highly complex” questions and unique data in life sciences, framing the collaboration as a way to potentially deliver medicines to patients faster if AI can help teams navigate that complexity. Moderna’s CEO Stéphane Bancel describes GPT-Rosalind as an important step in helping scientific teams reason across complicated biological evidence and workflows, which tracks with Moderna’s broader push to use machine learning across its mRNA platform.

From a business standpoint, GPT-Rosalind is also OpenAI’s entry into a biopharma vertical that is suddenly very crowded. Startups like Recursion and Insilico have been training biology-focused models for years, and Anthropic has already announced its own Claude for Life Sciences offering. Industry observers note that OpenAI is arriving a bit later than some rivals, but with a very different asset: a general family of GPT-5 models that already power a huge ecosystem, plus relationships with blue-chip enterprises far outside pharma. GPT-Rosalind fits neatly into that playbook – a verticalized layer on top of a general platform, with consulting giants like McKinsey, BCG, and Bain lined up as advisory partners to help big organizations identify use cases and integrate the model into their environments.

Not everyone is convinced this is a scientific revolution rather than a business move. Commentators have pointed out that, while the benchmark gains are real, they mostly show that Rosalind is better at particular narrow tasks than previous general models, not that it has cracked lab automation or experimental creativity. The fact that access is tightly restricted, with OpenAI selling a mix of model usage, infrastructure, and governance tooling to vetted enterprises, leads some critics to see GPT-Rosalind partly as an access product – a way to monetize high-stakes, high-margin customers who are willing to pay for specialized AI under strong controls.

There are also serious safety and dual-use concerns in the background. AI models capable of reasoning deeply about biology, designing sequences, or optimizing experiments could, if misused, lower barriers not just for beneficial research but for harmful biological designs. OpenAI is explicit that its trusted-access structure, governance requirements, and security controls are intended to mitigate this, and the company says it is working with national labs like Los Alamos to explore safe applications, including AI-guided protein and catalyst design that preserve or strengthen functional properties without enabling misuse. Still, the core tension remains: the more capable these systems become, the more powerful they are in both directions, which is why the current rollout feels more like a quiet pilot than a consumer product launch.

For scientists on the ground, the near-term impact of GPT-Rosalind will likely be pretty practical and somewhat unglamorous. Imagine dumping a pile of messy experimental outputs into a Codex project – a spreadsheet of dose-response curves, a log of cell viability assays, a PDF or two of prior studies – and asking the model to identify the same patterns an expert would, propose the next experiment, and link you to relevant literature you might have overlooked. Or handing it a protein of interest and having it walk you through potential functional domains, mutation effects, structural hypotheses, and analogous proteins in public datasets. The upside is time and cognitive bandwidth: fewer hours manually stitching together results and references, more time thinking about whether the biology really makes sense.

OpenAI is positioning this launch as the first step in a long-term commitment to “AI for science.” GPT-Rosalind is the debut entry in a broader life sciences model series, and the company says it plans to keep expanding the biochemical reasoning capabilities and support for long-horizon, tool-heavy workflows. The endgame vision is an increasingly capable partner that moves scientists faster from question to evidence, evidence to insight, and insight to new treatments. Whether that vision pans out will depend less on benchmark charts and more on what happens over the next few years in real labs: do projects move meaningfully faster, do failure rates drop, and do teams feel comfortable trusting an AI collaborator with parts of their reasoning loop?

For now, GPT-Rosalind is a clear signal that the era of science-first AI models has arrived. The big foundation-model players are no longer content to offer generic copilots and chatbots. They want a seat at the lab bench.