OpenAI just shipped GPT-Rosalind – its first biology-only model

The Hook — It Beat GPT-5.4 on BixBench

0.751 vs. 0.732. GPT-Rosalind outscored GPT-5.4 on BixBench, a comprehensive bioinformatics benchmark. Grok 4.2 (0.698) and Gemini 3.1 Pro (0.550) were further back.

The more striking result came from uncontaminated RNA sequence evaluation. OpenAI reports GPT-Rosalind's submissions ranked above the 95th percentile of human experts on sequence-to-function prediction, and at the 84th percentile on sequence generation.

Released yesterday (April 16), the model is called GPT-Rosalind – named after British chemist Rosalind Franklin, whose X-ray crystallography helped reveal DNA's double helix in the 1950s. Access isn't public. Not even ChatGPT Enterprise customers get it by default. Just nine hand-picked launch partners in a research preview.

What It Is — One-Sentence Definition + Three Differentiators

The definition: OpenAI's first frontier vertical-specialized model, tuned across biochemistry, protein engineering, and genomics to compress the evidence-synthesis, hypothesis-generation, and experimental-planning stages of life sciences research.

Three differentiators:

First, it's not general-purpose. Unlike GPT-5.4, "fundamental reasoning" is retuned for three domains: biochemistry, protein engineering, genomics. Molecular-level causal chains – where general LLMs hallucinate – hold up much better.

Second, tool use is a first-class citizen. Plug in PubMed search, AlphaFold structure prediction, molecular simulation runners, and GPT-Rosalind orchestrates evidence gathering → candidate modeling → result interpretation in a single loop.

Third, access is gated by the Trusted Access Program. OpenAI vets whether an organization is running legitimate health-improvement research and meets security/governance standards. Money alone doesn't get you in.

Source: marktechpost.com · press image, news citation

Core Specs — Head-to-Head vs. Competing Models

OpenAI didn't disclose parameter count, but benchmark results are public. Summary:

Benchmark	GPT-Rosalind	GPT-5.4	GPT-5	Grok 4.2	Gemini 3.1 Pro
BixBench (bioinformatics)	0.751	0.732	0.728	0.698	0.550
LABBench2 win rate vs GPT-5.4	6 of 11 task families	–	–	–	–
CloningQA delta	Biggest gain	Baseline	–	–	–
Uncontaminated RNA prediction	95th pct vs. human experts	Not reported	–	–	–
Uncontaminated RNA generation	84th pct vs. human experts	Not reported	–	–	–

LABBench2 is a new 2026 benchmark that spans 11 task families and roughly 1,900 biology research tasks. GPT-Rosalind beat GPT-5.4 on 6 of the 11, with the largest gap on gene cloning questions (CloningQA).

The "uncontaminated RNA" evaluation is closer to a real scientist workflow. It uses unpublished sequences to prevent data leakage – translation: "give it problems that aren't on the internet, and it still performs like a top-tier expert."

Feature Breakdown

Feature 1 — Four Research Loops

OpenAI names four target workflows:

• Evidence synthesis – scanning papers and datasets to map what's already been tested vs. what's still open
• Hypothesis generation – proposing the next experiment given current results
• Experimental planning – drafting protocols, control designs, risk checks
• Multi-step research tasks – chaining the three above into autonomous runs

OpenAI's blog explicitly frames the model as non-replacement: "the models won't replace scientists, but rather help them move faster through some of the most time-intensive and analytically demanding work." Compression, not substitution.

Feature 2 — Codex Life Sciences Plugin

GPT-Rosalind shipped with a sibling: the Codex Life Sciences plugin for GitHub.

Think of it as the biology-tuned cousin of GitHub Copilot. Researchers can invoke GPT-Rosalind from inside their own codebases – simulation scripts, analysis pipelines, lab automation code. The strategy: embed the model across the research pipeline, not just in a chat interface.

Source: onhealthcare.tech · news citation, fair use

Pricing + Availability — Access Is the Price

Item	Detail
Price	Undisclosed (enterprise contract)
Launch date	April 16, 2026
Access model	Trusted Access Program, vetted research preview
Region	U.S. enterprise customers only
Public API	Not available
ChatGPT Enterprise inclusion	Not included (separate product line)
Eligibility	Legitimate health-improvement research + security/governance review

The nine launch partners, per OpenAI's blog: Amgen, Moderna, Allen Institute, Thermo Fisher Scientific, Oracle Health and Life Sciences, NVIDIA, Benchling, and UCSF School of Pharmacy.

Who It's For — Personas + Alternatives

Persona	Value	Alternative
Big Pharma R&D	Hypothesis screening speed	Isomorphic Labs AlphaFold 3, Insilico Medicine
mRNA platform company	Automated vaccine candidate design	In-house models + DeepMind partnerships
Nonprofit research institute	High-precision analysis of public datasets	Hugging Face BioGPT, academic open models
Lab equipment / reagent vendor	Workflow AI integration	Med-PaLM API, Claude for Life Sciences

The partner list tells a single story. OpenAI wants both ends of the research pipeline – the front end (planning, hypothesis) and the back end (instruments, data). Thermo Fisher is the "lab OS" layer scientists already use, and Benchling is a bio R&D data platform. With Oracle Health and NVIDIA attached, you can bundle storage + compute + model in one stack.

Competitive Response + Market Position

The life sciences AI landscape was Google's territory first.

DeepMind shipped AlphaFold 3 in 2024, extending structure prediction to protein-ligand interactions. Isomorphic Labs (Alphabet subsidiary) locked in ~$600M-scale partnerships with Novartis and Eli Lilly in 2024–2025. Gemini for Biology and Med-Gemini are already in clinical partner pipelines.

That makes OpenAI's timing pointed.

The era of "one general GPT API covers every vertical" is ending. GPT-Rosalind is the signal that OpenAI is formally opening a domain-specialized product line.

The market reacted the day of launch. April 16, shares of AI-bio middle players Recursion, Schrödinger, and Absci all closed lower. The question every board is asking: "If OpenAI walks directly into our customer set, what's our defensibility?"

What Changes for Whom

For pharma companies: The bottleneck shifts from analysis to validation. If AI handles evidence synthesis and hypothesis generation, the real gating factor becomes wet-lab verification throughput. Pipelines can shorten 1–2 years, but only if internal lab automation catches up.

For bio startups: Positioning reset time. OpenAI just claimed the "foundational reasoning" layer. Survival now means either a disease-area specialty (rare oncology, neurodegeneration) or a physical-world hook – lab automation, robotic assays, proprietary reagents.

For general developers and users: You're not going to use GPT-Rosalind directly. But the signal matters. When domain-specialized GPTs become a product line, expect legal, financial, and materials-science variants within 12 months. The one-API-fits-all-verticals era is closing.

References

• OpenAI: Introducing GPT-Rosalind
• Axios: OpenAI launches new AI model for life sciences research
• Bloomberg: OpenAI Takes on Google With New AI Model
• VentureBeat: GPT-Rosalind limited access preview
• MarkTechPost: GPT-Rosalind built for drug discovery and genomics research
• Euronews: What to know about OpenAI's new life sciences model