OpenAI opens GPT-Rosalind to life sciences teams

OpenAI has moved GPT-Rosalind beyond its research preview, giving eligible organizations access through the API, Codex, and ChatGPT Enterprise. Named after Rosalind Franklin, the model targets multi-step work in early drug discovery, genomics, protein engineering, and translational medicine.

The release combines a specialized reasoning model with a public research plugin for Codex. The plugin connects supported models to more than 50 scientific tools, databases, and literature sources, providing an orchestration layer for workflows that span multiple systems.

Fragmented research gets a reasoning layer

Developing a drug from target discovery to U.S. regulatory approval often takes 10 to 15 years. Researchers must reconcile published literature, proprietary results, sequence databases, experimental protocols, and changing hypotheses, frequently through manual searches and disconnected software.

GPT-Rosalind is designed to synthesize evidence, generate hypotheses, plan experiments, and carry context across those steps. A typical task might involve reviewing papers and datasets for a biological target, assessing conflicting evidence, proposing follow-up experiments, and documenting the analysis for collaborators.

Benchmarks favor Rosalind, with caveats

OpenAI evaluated GPT-Rosalind on two public benchmark suites and an unpublished industry task. The reported results cover literature retrieval, database access, sequence analysis, protocol design, prediction, and sequence generation.

Evaluation	Task	Reported result
BixBench	Real-world bioinformatics and data analysis	Leading performance among models with published scores
LABBench2	Literature retrieval, database access, sequence manipulation, and protocol design	Outperformed GPT-5.4 on 6 of 11 tasks
CloningQA	End-to-end design of DNA and enzyme reagents for molecular cloning	Produced the largest reported gain within LABBench2
Dyno Therapeutics	RNA sequence-to-function prediction and sequence generation	Best-of-10 submissions ranked above the 95th percentile for prediction and near the 84th percentile for generation

The Dyno Therapeutics evaluation compared Codex submissions with 57 historical scores from human experts in AI and biology. It used held-out sequences intended to reduce training-data contamination, but the task remains unpublished and cannot yet be independently reproduced. The best-of-10 method also measures the strongest result across repeated attempts, so developers should not treat it as equivalent to single-run performance.

The plugin carries the integration load

The Codex plugin packages modular skills for human genetics, functional genomics, protein structure, biochemistry, clinical evidence, and public-study discovery. It can coordinate sequence searches, structure lookups, literature reviews, and dataset discovery across more than 50 public multi-omics databases and scientific tools.

Eligible Enterprise customers can pair the plugin with GPT-Rosalind for specialized biological reasoning. Other users can apply the same package to OpenAI’s mainline models, allowing teams to inspect and test the orchestration layer before receiving access to Rosalind.

OpenAI has not provided endpoint names, rate limits, context limits, or production pricing in the announcement. Teams planning an integration will need those details, along with the plugin’s individual data-source requirements, before estimating latency, cost, and operational dependencies.

Access starts behind a governance gate

GPT-Rosalind uses a trusted-access deployment model for qualified Enterprise customers. Applicants must demonstrate a legitimate research purpose, organizational governance, access controls, and a secure operating environment.

OpenAI describes the broader release as available to eligible organizations worldwide, while the initial trusted-access rollout begins with qualified U.S. Enterprise customers. Organizations outside the United States should confirm regional availability during onboarding.

Usage currently does not consume existing credits or tokens, subject to abuse guardrails. OpenAI characterizes these terms as preview pricing and has not announced what the model will cost as access expands.

Early participants include Amgen, Novo Nordisk, Moderna, Thermo Fisher Scientific, Oracle Health and Life Sciences, NVIDIA, the Allen Institute, Benchling, and the UCSF School of Pharmacy.

Workflows suited to the model

OpenAI’s evaluations and product description position GPT-Rosalind for several recurring research tasks:

1. Reasoning across molecules, proteins, genes, pathways, and disease biology within one investigation
2. Interpreting sequence-to-function relationships for DNA, RNA, and proteins
3. Designing molecular cloning protocols, including reagent selection
4. Combining published research with internal experimental results to assess a biological target
5. Planning follow-up experiments and generating quality-control reports or interactive notebooks in Codex

The model supports reasoning and software orchestration around wet-lab work, while experimental validation remains necessary. Dyno’s reported gap between prediction and sequence-generation performance also indicates that proposing novel biological designs remains harder than scoring existing candidates.

A two-layer release strategy

OpenAI is distributing specialized biological reasoning and research infrastructure through separate channels. GPT-Rosalind remains behind eligibility and governance controls, while the public plugin gives a wider group of developers access to the associated tool-use workflows with mainline models.

This structure lets research teams begin evaluating database connections, workflow design, provenance, and human-review requirements without waiting for model approval. Teams handling proprietary or sensitive data must still assess each external tool’s retention, security, licensing, and compliance terms before using the plugin in production.