OpenAI Limits Rosalind Subsidy to U.S. University Labs

OpenAI is offering eligible United States university labs up to 30 no-cost seats for its GPT-Rosalind life-sciences AI for up to a year, but approval is required and the plan excludes API access.

Announced on October 15, the plan expands access to an existing model, not evidence that AI is producing better medicines. The test is whether researchers can improve decisions beyond company-run benchmarks.

What academic labs receive

A principal investigator or designated lab lead applies for the whole team. Approved labs can receive up to 30 ChatGPT Business Standard seats at no cost for up to 12 months, with GPT-Rosalind access activated only after a separate trusted-access review. Other commercially available frontier OpenAI models are included subject to plan usage limits.

The offer is not a research-computing grant: it includes no API access or credits for automated pipelines. OpenAI says a payment card is required at setup and advises labs to check the offer’s end date and any charges; seats are advertised at $0 per user per month during the offer.

• Eligibility: academic research labs at United States universities; a PI or lab lead applies.
• Scale: up to 30 named members in one workspace for up to 12 months.
• Model access: GPT-Rosalind requires trusted-access approval and activation.
• Not included: API access or API credits; institutions with an Enterprise agreement use that route.

OpenAI says research data shared in the plan is “not used to train our models by default.” Its help centre also says standard Business data-retention rules apply, an administrator may be able to access or manage workspace data, and Zero Data Retention is not available under the offer. Labs therefore need to check university policy and data controls before uploading sensitive research.

A research assistant, not a drug pipeline

GPT-Rosalind is OpenAI’s life-sciences model family for literature review, medicinal chemistry, genomics and experimental planning. Its Workbench connects model reasoning with specialist tools. Since September 11, eligible organizations globally have been able to request it through a separate trusted-access route, not this subsidy.

OpenAI reports 27.5% on MedChemBench for Rosalind against 25.1% for GPT-5.5, using 7.2% fewer tokens. On GeneBench, it reports 21.6% accuracy against 20.4%, with 31% fewer tokens. These defined medicinal-chemistry and genomics evaluations are not clinical trials or proof a drug candidate will succeed.

That gap between benchmark performance and patient benefit remains a central test for the field. Andreas Bender, co-author of a 2026 peer-reviewed Perspective on AI in drug discovery, wrote that “evidence of their clinically relevant impact is, so far, disappointingly limited.” The paper cites challenges in translating models to clinical goals, working with conditional biological data, defining real-world problems and integrating tools into research practice.

Access controls and safety trade-offs

The plan requires institutional and PI verification, a description of the lab’s intended research, review of its activities, and a named institutional biosafety contact. The lab lead controls which members receive access. OpenAI says its safety monitoring and enforcement continue after activation; approval can be refused or revoked.

The GPT-Rosalind-5.5 system card, published in June before this plan, describes model boundaries and post-access monitoring using automated safety classifiers, account- or organization-level indicators and specialist review when available. But it says OpenAI does not use automated monitors to block potentially unsafe generations in real time. Other review measures exist; the card does not describe an automatic filter on every answer.

The same system card reports that Rosalind performed significantly below a freely available gradient-based method on one DNA-sequence-design task. The result narrows any claim of universal superiority: scientific value will depend on the task, independent validation and researchers’ ability to detect errors.

A U.S. subsidy, with a global-access question

Only labs at U.S. universities can apply for the no-cost academic plan. A research group in Nairobi, Lagos, Dakar or another non-U.S. institution is outside its eligibility, even though eligible organizations globally can seek Rosalind through the separate business or enterprise approval path. OpenAI says it plans to expand academic access to other countries in the future, but has not announced a timetable.

Steven Wright’s article, published online on March 12, 2026, in the Journal of English for Research Publication Purposes, argues that unequal AI and computing access may create a two-tier system of knowledge production, drawing on studies from Africa, Asia and Latin America. It does not evaluate Rosalind, but raises a relevant question about a U.S.-only subsidy.

What evidence should come next?

OpenAI lists UCSF’s Zhao Lab, Mathieu Bourdenx’s lab and Brian Hie’s Stanford/Arc Institute lab as exploring Rosalind, but gives no aggregate counts of applications, approvals, projects or outcomes. The test is whether labs can show a reproducible improvement over existing workflows.

• How many labs apply, qualify and activate seats—and what happens when the 12-month period ends?
• Do users publish reproducible work or experimentally validate model-generated hypotheses?
• How are data access, biosafety review, and erroneous outputs handled in actual projects?
• Will the subsidized plan expand beyond the United States or add API access?

For now, the plan is a limited experiment in making a specialist research assistant cheaper to try. Its scientific case will rest on results that labs can validate and reproduce, while its access case will depend on whether those opportunities reach researchers beyond the U.S. institutions eligible today.