OpenAI's GPT-Rosalind Targets Drug Discovery, Beats GPT-5.4

OpenAI launched GPT-Rosalind on April 17 — its first domain-specific model, built explicitly for biology, genomics, and drug discovery. On the bioinformatics benchmarks the company published, it outperforms GPT-5.4, Grok 4.2, and Gemini 3.1 Pro on the tasks that matter to working scientists.

The model is named after Rosalind Franklin, the biophysicist whose X-ray diffraction images revealed the structure of DNA. It's a pointed choice: Franklin's work went underrecognized for decades, and OpenAI is positioning Rosalind as a tool built to actually credit and accelerate the kind of research she did.

Why this matters: Until now, frontier AI models have been generalists that happen to do some biology well. GPT-Rosalind is the first major attempt by a foundation model lab to build a vertical model for science. If it works at the bench, it signals a new era: purpose-built AI for individual research disciplines, not one model trying to do everything.

What Rosalind Does

GPT-Rosalind is designed for multi-step scientific workflows — the real work of research, not just answering questions. OpenAI describes four core capabilities: evidence synthesis across scientific literature, biological hypothesis generation, experimental planning, and sequence-to-function analysis for proteins, genes, and RNA.

The model ships with a companion Life Sciences research plugin for Codex that connects to more than 50 specialized tools and databases. Those include human genetics databases, functional genomics repositories, protein structure prediction systems, multiomics catalogs, and clinical evidence libraries. Researchers don't have to manually bounce between BLAST, UniProt, AlphaFold results, and ClinicalTrials.gov — Rosalind orchestrates the workflow.

Crucially, the model was evaluated on held-out, unpublished RNA sequences to prevent benchmark contamination. On sequence-to-function prediction, its best-of-ten submissions ranked above the 95th percentile of human experts. On sequence generation, it hit roughly the 84th percentile. Those are working-scientist numbers.

The Benchmark Story

On BixBench — a bioinformatics evaluation built by Edison Scientific to stress-test real computational biology tasks — GPT-Rosalind scored 0.751 pass@1. GPT-5.4 hit 0.732. GPT-5 managed 0.728. Grok 4.2 scored 0.698. Gemini 3.1 Pro came in at 0.550.

On LABBench2, a 2026 benchmark spanning roughly 1,900 biology research tasks across eleven categories, Rosalind beat GPT-5.4 on six of eleven task families. The widest margin was on CloningQA, which tests end-to-end molecular cloning reagent design — one of the most procedurally complex tasks in a wet lab.

The gap over GPT-5.4 isn't enormous, but it's meaningful. In practice, a bioinformatics pipeline that gets 75% of tasks right instead of 73% is running noticeably fewer broken experiments. And the margin expands on the hardest, most domain-specific tasks — exactly where a specialized model should win.

Who Gets Access

GPT-Rosalind is not publicly available. OpenAI is routing it through a trusted-access program limited to vetted enterprise customers in the United States. Launch partners include Amgen, Moderna, Thermo Fisher Scientific, the Allen Institute, Oracle Health and Life Sciences, NVIDIA, Benchling, and UCSF School of Pharmacy.

The rationale is dual-use risk. A model that can design RNA sequences or predict protein function is also — in theory — a model that could help someone design a pathogen. OpenAI is explicitly gating access to reduce that risk, with partner organizations agreeing to usage controls, auditing, and safety review protocols.

For researchers outside the trusted-access program, the free Codex Life Sciences plugin offers a partial substitute. It connects GPT-5.4 to the same 50+ scientific databases, which is genuinely useful even without Rosalind's specialized reasoning on top.

Industry Impact

The drug discovery AI landscape just got more crowded. Google DeepMind has AlphaFold and the newer AlphaGenome. Isomorphic Labs — spun out of DeepMind — is running a full pharma pipeline. Recursion, Insitro, and Absci are all building AI-native drug discovery platforms.

OpenAI's entry is different in one important way: Rosalind isn't designed to replace the specialized tools. It's designed to orchestrate them. A researcher can stay in a conversational interface while the model handles literature review, protein structure lookup, sequence design, and experimental planning across multiple systems.

That positioning threatens Google's vertical integration play less directly than it threatens the fragmented tooling ecosystem that most biology labs currently stitch together manually. If Rosalind works as advertised, the question for Benchling, DNAnexus, and dozens of smaller vendors is whether they become features of OpenAI's platform or find defensible positions outside it.

What's Next

Launch partners are already running pilot programs. Moderna has said it will evaluate Rosalind against its internal AI tools for target identification and preclinical candidate selection. Amgen is testing it for biomarker discovery workflows. Thermo Fisher is exploring integration with its instrument automation platforms.

OpenAI hasn't committed to a timeline for broader availability, but the pattern from similar trusted-access launches suggests a wider rollout within 6-12 months, assuming the pilots clear safety review. For academic researchers, a tiered access program with institutional approvals seems likely.

The Bottom Line

GPT-Rosalind is the first credible answer to the question: what does a frontier model look like when you build it for one thing instead of everything? If it holds up in actual discovery workflows — not just benchmarks — expect more vertical models to follow. If you work in pharma, biotech, or academic life sciences, this is the product to watch over the next year.