OpenAI's GPT-Rosalind ships through Codex, and the plugin is doing the real work

Three front doors, one gated, two open

OpenAI is releasing GPT-Rosalind, a reasoning model aimed at biology, drug discovery, and translational medicine, as a research preview in ChatGPT, Codex, and the API. Access is limited to qualified customers through a trusted-access program, starting with U.S. Enterprise organizations.

Alongside the model, OpenAI shipped a Life Sciences research plugin for Codex, available today on GitHub. The split is deliberate: the model itself is gated, but the plugin is available more broadly, and the announcement notes that all users can run the plugin package against OpenAI's mainline models — not just GPT-Rosalind.

That means the frontend surface — the thing a scientist actually opens and types into — is decoupled from the frontier model behind it. You can have the connective tissue without the flagship reasoning, or both together if you clear the eligibility bar.

The plugin is the orchestration layer scientists touch

OpenAI describes the plugin as "an orchestration layer" of modular skills spanning human genetics, functional genomics, protein structure, biochemistry, clinical evidence, and public study discovery. It exposes access to more than 50 public multi-omics databases, literature sources, and biology tools.

The framing matters. The announcement repeatedly argues that life-sciences progress is constrained not just by hard science but by "fragmented" research workflows — literature, specialized databases, experimental data, evolving hypotheses. The plugin is OpenAI's answer to that fragmentation, and it ships as skills for "common repeatable workflows such as protein structure lookup, sequence search, literature review, and public dataset discovery."

In other words, the durable product here may be less the model weights and more the standardized skill package that turns a chat interface into a tool-calling research assistant. The model swaps out; the workflow surface stays.

An evaluation run inside the app, not just the model

The most concrete performance claim is a partnership with Dyno Therapeutics on an RNA sequence-to-function prediction and generation task, using unpublished, uncontaminated sequences and compared against 57 historical human-expert scores.

Read the method carefully: submissions were "evaluated directly in the Codex app," and the result cited is "best-of-ten model submissions" ranking above the 95th percentile of human experts on prediction and around the 84th percentile on generation. That is a best-of-ten figure produced through the app frontend, not a single-shot model score — a distinction worth holding onto when comparing to a lone human expert's one attempt.

On public benchmarks, the claims are narrower and more precise: leading performance on BixBench among models with published scores, and beating GPT-5.4 on 6 of 11 LABBench2 tasks, with the standout being CloningQA's end-to-end DNA and enzyme reagent design.

Who gets in, and who signs off

The trusted-access structure is the other half of the frontend story. To use the model, organizations must be doing legitimate research with clear public benefit, maintain governance and misuse-prevention controls, and restrict access to approved users in secure environments. OpenAI names Amgen, Moderna, the Allen Institute, and Thermo Fisher as customers, and Los Alamos National Laboratory for AI-guided protein and catalyst design.

The life sciences field demands precision at every step. The questions are highly complex, the data are highly unique, and the stakes are incredibly high. — Sean Bruich, Senior Vice President of Artificial Intelligence and Data, AmgenMontana Labs

Notably, integration help comes not from OpenAI alone but from advisory partners McKinsey, BCG, and Bain, positioned to "identify high-impact use cases" and "integrate the model into enterprise environments." That signals the onboarding surface is a consulting-mediated enterprise deployment, not a self-serve sign-up.

The takeaway: build against the plugin, not the model name

For teams watching this space, the actionable detail is that the Codex plugin is open on GitHub and works with mainline models today, while GPT-Rosalind sits behind qualification and safety review. The reusable investment is the tool-connected workflow, not access to any one model.

OpenAI is explicit that this is the first release in a series and that it will keep swapping in more capable models "as the workflows themselves become more complex." If you standardize your research pipeline on the plugin's skills and 50-plus connectors now, you inherit model upgrades without rebuilding the interface — which is exactly the bet OpenAI seems to be making by shipping the frontend more freely than the frontier.