On this page
Why it matters here
OpenAI's October 30 addendum reports higher scores for GPT-Rosalind-6 than GPT-Rosalind-5.5 on three life-science research benchmarks using refreshed common problem sets. The report distinguishes medicinal-chemistry rubric scores, single-attempt biological-analysis success and applied research-task scores, while disclosing task exclusions and a non-production GeneBench configuration. Rosalind-6 uses Astra's weights without additional training. This supplies qualified context about reported scientific-model capability, rather than a measured research-acceleration or recursive-improvement result.
Research autonomyEvaluation integrity
What to keep in mind
- Vendor report, not independent experimental replication. The reported comparisons use refreshed common problem sets; earlier Rosalind launch scores are not interchangeable.
- MedChemBench v7 uses 99 problems and up to nine scored responses per problem; GeneBench Pro v13 estimates single-attempt success from seven responses on 129 problems; LifeSciBench Gold excludes six flagged problems and evaluates the remaining 94 with up to eight responses per problem. Response scores are averaged within problems and problems weighted equally.
- Reported beneficial scores use the highest reasoning budget tested. Exact reasoning token budgets, full agent harnesses, grader identities/rubrics, item-level results, all attempted-response/failure accounting and uncertainty are not supplied in the checked beneficial-evaluation section. GeneBench uses a benchmark-specific configuration that differs from production.
- GPT-Rosalind-6 uses GPT-6 Astra's weights without additional training or post-training. Reduced refusals and deployment configuration must not be mistaken for a separately trained biological model or a controlled improvement over Astra.
- Human researchers contributed LifeSciBench tasks; this is not a matched human-performance comparison or evidence of time saved in research. The beneficial tests differ from the card's safety-capability evaluations.
- The source inherits Astra's AI Self-Improvement assessment; this entry does not establish retained recursive improvement, an integrated autonomous research loop, a cross-model ranking or a tracker autonomy assessment. No numeric observation, benchmark protocol, public figure or homepage synthesis is proposed.
- Checked original HTML and PDF Sections 1–3, including all three beneficial-evaluation sections; PDF page 5 (printed 4) visually inspected. Underlying benchmark datasets, full item-level records and parent-card experiments were not independently audited.