OpenAI opens GPT-Rosalind-⁠6, which refuses fewer biology questions, to approved researchers

Same weights, different refusals

In a system card addendum published Oct. 30 on its Deployment Safety Hub, OpenAI described GPT-Rosalind-6 as a frontier reasoning model for biology, drug discovery and translational medicine. It is offered to qualified scientists, research institutes, companies and government partners as GPT-Rosalind-6 Research; the help center calls the product GPT-6 Rosalind Research, and the API model ID is gpt-6-rosalind. OpenAI has not published a separate blog announcement for Rosalind-6.

According to the system card, GPT-Rosalind-6 uses the same weights as GPT-6 Astra, OpenAI’s flagship reasoning model; no training data was added and no additional post-training was done for this release. The difference lies in behavior and access: the model has fewer refusals on dual-use biology queries but still refuses requests for biological weaponization and retains refusals on other safety topics. That is why it is only made available through trusted access.

A biorisk knowledge test built with Gryphon Scientific puts numbers on the difference: the refusal rate is 9.4% for GPT-Rosalind-6, 29.2% for GPT-6 Astra and 1.8% for the previous GPT-Rosalind-5.5. On the same test, accuracy measured by consensus across 32 responses is 78.3% for Rosalind-6, 79.2% for Rosalind-5.5 and 63.3% for Astra.

OpenAI’s own benchmarks

OpenAI compared the model with GPT-Rosalind-5.5 on three life sciences benchmarks; the 5.5 scores were re-run on the same benchmark versions and problem sets, so they differ from those reported at its launch. At the highest reasoning budget, on the 99 problems of MedChemBench v7, which tests chemical structure interpretation, structure-activity relationships and predictions of drug potency and toxicity, the mean score rose from 28.9% to 49.6%. On the 129 problems of GeneBench Pro v13, which tests planning and running quantitative analyses of biological data, single-attempt success rose from 16.7% to 37.7%; that evaluation used a benchmark-specific configuration that differs from production. On the 94 problems of LifeSciBench Gold, built from tasks contributed by practicing life scientists, the score rose from 44.0% to 61.3%. All of these are OpenAI’s own measurements.

Under its Preparedness Framework, OpenAI treats Rosalind-6 as High capability in the biological and chemical domain, short of the Critical threshold. It relies on GPT-6 Astra’s assessments for cybersecurity, where it treats the model as Critical, and for AI self-improvement, where the model does not reach High. On the biology side, after an adjustment that counts refusals as correct answers, the model exceeds the indicative threshold on three of the four High capability evaluations; none of the four Critical capability evaluations met its threshold.

Access and pricing

The model is available only to approved customers. OpenAI reviews each applicant’s proposed use, security controls and governance practices, and runs business verification and compliance screening for non-government applicants. There is a separate pathway for government entities working on biosecurity, biosafety, public-health preparedness and biodefense. According to the system card, this release does not broaden eligibility beyond prior announcements such as the Rosalind Academic Plan. Customers must apply qualified human review before model outputs feed into experiments, publications or significant decisions. The model also keeps GPT-6 Astra’s standard cyber safeguards, including real-time blocking.

According to the help center, GPT-6 Rosalind Research is available to a limited set of approved Research organizations in Work and Codex, and through the API where separately approved; it is not available in ChatGPT’s Chat tab. In Enterprise workspaces it starts in a two-week Admin Preview and is off by default during that period. Rosalind has two access tiers: Research covers GPT-6 Rosalind Research and GPT-5.5 Rosalind Research, while Discovery covers Rosalind 5.5 with stricter monitoring.

According to OpenAI’s API pricing page, gpt-6-rosalind costs $15 per million input tokens and $75 per million output tokens. Cached input is $1.50 per million tokens and cache writes are $18.75; long-context prices are $30 and $112.50. The Rosalind series was announced as a research preview on April 16, 2026, and moved out of preview on Sept. 11, opening to eligible organizations through trusted access.