OpenAI has released a system card addendum for the GPT-Rosalind-6 model, detailing deployment safety measures for its advanced capabilities. The model uses the same weights as GPT-6 Astra, with no additional training data or post-training applied. In genomics and quantitative biology, GPT-Rosalind-6 achieved a 37.7% single-attempt success rate on the Gene Bench Pro v13 benchmark at the highest reasoning budget, compared to 16.7% for GPT-Rosalind-5.5.
OpenAI classifies the launch as High capability in the Biological and Chemical domain, employing four High capability and four Critical capability evaluations. These tests assess wet lab knowledge, troubleshooting, and biological design capabilities that could potentially enable dangerous novel threats. To support legitimate advanced biological research, GPT-Rosalind-6 is designed to be more useful than generally available Chat GPT models, though it remains tuned to refuse or safely respond to requests enabling severe biological weaponization.
The model operates on a trusted-access deployment basis. Safety evaluations track model responses to held-out synthetic data, red-teaming, and production data, with benchmarks updated continuously. Biosecurity - Severe benchmark results are prioritized for trusted-access deployments, while Biosecurity - Benign evaluates model compliance on operational questions.
Pass@1 measures single-attempt success, with refusal-adjusted scores estimating capabilities safety training may conceal.
Source: Hacker News · Summarized by HeadlinesBriefing