OpenAI's GPT-Rosalind-6 brings its flagship GPT-6 Astra to vetted biology labs, with fewer refusals

GPT-Rosalind-6 isn't a new model under the hood. It uses the same weights as GPT-6 Astra, OpenAI's flagship reasoning model, and OpenAI says it added no training data and did no further post-training for the release[2].

The difference is how it behaves and who gets it. Rosalind refuses fewer dual-use biology questions than generally available ChatGPT models, so it can help with legitimate research that GPT-6 Astra might decline in general access. It still refuses requests that would meaningfully help with biological weapons, and keeps refusals on other safety topics[3].

OpenAI compared the model with its predecessor, GPT-Rosalind-5.5, on three life sciences benchmarks using the same problem sets, reporting results at the highest reasoning budget it tested[3]:

• MedChemBench v7 (medicinal chemistry, 99 problems): 49.6% mean rubric score, up from 28.9%.
• GeneBench Pro v13 (genomics and quantitative biology, 129 problems): 37.7% single-attempt success, up from 16.7%. OpenAI notes this run used a benchmark-specific configuration that differs from production.
• LifeSciBench Gold (tasks from practising life scientists, 94 problems after six were excluded): 61.3%, up from 44.0%.

Under its Preparedness Framework, OpenAI treats GPT-Rosalind-6 as High capability in the biological and chemical domain, below Critical. After adjusting for refusals, it exceeded the indicative High thresholds on three of four wet-lab knowledge and troubleshooting evaluations, while none of the four Critical-level tests of biological design, such as protein function and virus cell-entry prediction, met their thresholds[3].

For other risk areas it relies on GPT-6 Astra's assessments, so Rosalind is treated as Critical in cybersecurity and below High in AI self-improvement. It keeps Astra's real-time cyber blocking and adds a more conservative cyber refusal boundary[1].

Access follows the rules already used for GPT-Rosalind-5.5, and OpenAI says this release doesn't broaden eligibility. It reviews whether applicants have a legitimate life sciences mission, a credible use case, sound governance and the ability to control access. Non-government applicants also go through business verification and compliance screening, and there's a separate pathway for government biosecurity, public-health and biodefence work[3].

Approved customers must limit use to authorised staff with least-privilege controls, and have qualified people review model outputs before they inform experiments, publications or other consequential decisions. OpenAI says it can restrict or revoke access for misuse and is repeating a biosecurity red-teaming exercise on the new model now that it's deployed[3].

Every number here comes from OpenAI's own system card, with no independent replication. The document doesn't cover pricing, approval timelines or how many organisations will qualify[4].

Rosalind shows where frontier labs are heading with sensitive domains: instead of training a separate specialist model, OpenAI ships the same flagship weights with a looser refusal policy behind a gate of vetting, monitoring and contracts. For biotech and pharma teams, the model's capability now comes bundled with a compliance process, and getting through that process becomes part of adopting it.