GPT‑Rosalind Exits Research Preview and Expands Globally, with Codex Plugin Connecting More Than 50 Life Sciences Data Sources

OpenAI updated the GPT‑Rosalind launch page on September 11, announcing that its life sciences–specific model was exiting research preview and would be made available to eligible institutions worldwide through a trusted-access program. Approved customers can use it in ChatGPT, Codex, and the API, and will automatically receive access to the latest Rosalind models going forward. This is not a general release through the standard API: applicants must still demonstrate that their research serves the public interest, maintain governance and misuse safeguards, and restrict access to approved personnel. [OpenAI’s update](https://openai.com/index/introducing-gpt-rosalind/) also states that standard pricing will take effect on October 5.

Of greater engineering significance is the simultaneous release of the [Life Science Research Plugin](https://github.com/openai/plugins/blob/main/plugins/life-science-research/README.md). Rather than serving as a connector to a single database, it is an orchestration layer comprising 50 skills. By default, `research-router-skill` first identifies the question type, normalizes entities such as genes, proteins, diseases, variants, compounds, and accessions, and then selects the minimum set of required skills. Independent evidence paths can be searched in parallel, after which a coordinating agent cross-checks the results. Its coverage includes human genetics, functional genomics, expression data, protein structures, chemistry, clinical evidence, and public research data, with connections to more than 50 tools and data sources. The plugin itself can also be used with mainstream models and does not require Rosalind access.

The performance figures should still be interpreted cautiously. OpenAI says Rosalind outperformed GPT‑5.4 on six of LABBench2’s 11 tasks. In tests conducted with Dyno Therapeutics using non-public RNA data, a best-of-ten approach placed its predictions and sequence generation above the 95th and 84th percentiles of human experts, respectively. Selecting the best result from multiple attempts increases inference costs, while the use of non-public data prevents independent reproduction. The existing [5.5 system card](https://deploymentsafety.openai.com/gpt-rosalind-5-5) separately rates the model’s biological and chemical capabilities as High but below Critical. Engineering teams should next verify the actual API model identifier, data residency, tool permissions, traceable citations, and the total cost of multi-step research workflows under standard pricing.