Model Card | GPT-6 Rosalind

🌏 中文版

Model Info

Highlights

• Uses exactly the same weights as GPT-6 Astra. OpenAI states plainly that it "did not add training data or perform additional post-training for this release" — this launch changes safety tuning and the access gate, not model capability
• All three published life-sciences benchmarks more than doubled versus predecessor GPT-Rosalind-5.5: MedChemBench v7 (medicinal chemistry) 28.9% → 49.6%, GeneBench Pro v13 (quantitative biology) 16.7% → 37.7%, LifeSciBench Gold (life sciences research) 44.0% → 61.3%
• On the biosecurity compliance suite, it fully answers "benign biology" questions 100% of the time (Astra: 91.6%) — but that gain comes mostly from refusing less, not from answering more correctly
• Ships alongside Rosalind Workbench, a research workspace inside Codex/ChatGPT with built-in next-generation sequencing workflows: FASTQ quality control, bulk RNA-seq, and single-cell RNA-seq

Benchmark Results

⚠️ All figures are OpenAI's own self-reported results from the system card addendum; none have been independently reproduced. GeneBench Pro used "a benchmark-specific configuration that differs from production." The refusal-adjusted biorisk score treats a refusal as if it were a correct answer, to estimate the true capability that safety training may be hiding — it's a risk-assessment tool, not a general capability ranking.

Versus Predecessor / Competitors

Compared with GPT-Rosalind-5.5, the more-than-doubled scores on all three life-sciences benchmarks come mainly from the underlying model moving from GPT-5.5 to GPT-6 Astra, not from anything new in this release itself — in other words, this card is really quantifying how much stronger GPT-6 Astra is than GPT-5.5, repackaged as progress in the life-sciences series.

Compared with Astra on the same weights, Rosalind leads on the unadjusted biorisk knowledge test (78.3% vs 63.3%) — but mainly because its refusal rate is only 9.4% against Astra's 29.2%. Once refusals are treated as "the model knew the answer but chose not to give it" and scores are re-estimated, the ranking flips: the addendum reports Rosalind at 85.0% and Astra at 90.0%, meaning Astra's adjusted capability is actually higher. i-SCOOP's read is blunt: the apparent advantage "comes from answering more often, not from answering better." On the sensitive-biosecurity safety grader, Rosalind only passes 79.0% of the time — 20.8 points below Astra — trading more answered questions for a higher misclassification risk.

The pricing strategy is the part worth remembering. OpenAI has started tiering the price of the same weights by safety configuration and access program: general-access gpt-6-astra runs $10/$50, trusted-access gpt-6-rosalind with a looser refusal boundary runs $15/$75 (1.5x), and the pricing page also lists a cheaper gpt-rosalind-research at $5/$25 that OpenAI hasn't clearly mapped to a specific version. Run the numbers on a research-agent task of 200K input tokens and 20K output tokens: gpt-rosalind-research costs $1.50, Astra costs $3.00, and Rosalind 6 costs $4.50 — at 10,000 runs a month that's $15,000 versus $30,000 versus $45,000. Anthropic entering the same market with Claude Science suggests "charging a premium for a lower refusal rate" could become a standard pricing pattern for life-sciences AI in the second half of 2026.

What This Means for Agent Development

The most direct impact on agent architecture is that "human review" has moved from a soft recommendation to a contractual obligation: the system card addendum explicitly requires customers to "apply qualified human review before model outputs inform experiments, publications, or consequential decisions." If you're building a research agent that proposes compounds or queues experiments automatically, your closed-loop design must include a named review checkpoint with a retained record — you can't just let the model decide on its own. OpenAI doesn't define what counts as "qualified" review or what a reviewer must sign off on, so that part is on you to spec out.

• If you already have trusted-access approval and the governance/access-control infrastructure to go with it (say, a GxP-validated environment) at a life-sciences research organization: Rosalind is worth evaluating. Astra still refuses or caveats roughly 8% of benign biology prompts, which is a real interruption risk for an agent pipeline meant to run unattended — Rosalind pushes that close to zero
• If your agent reads external or untrusted content (pulling from literature databases, collaboration messages, and the like): note that one metric in the addendum got notably worse — "unintended engagement with external agent messages" rose from 0.0% on Astra to 2.5% on Rosalind. That's the direct cost of loosening the refusal boundary, and it's a real security regression for any tool-using agent that consumes external input
• Not a fit: teams that haven't received trusted-access approval yet (you can't even apply the model without it); customer-facing products or CRO-for-hire offerings (the addendum explicitly restricts use to "approved internal research"); teams without a revocation plan or fallback model — OpenAI reserves the right to pull access at any time for "misuse" but gives no stated triggers or notice period, so a pipeline with a single point of failure on this model has no safety net

Today's Takeaway

I used to assume "new version" meant "new training" meant "better capability." OpenAI spells out the opposite here: GPT-6 Rosalind added no new training data or post-training, and nearly all of its benchmark gains trace back to the stronger Astra base underneath it. What this release actually changes — and what it actually charges for — is how the same set of weights behaves under different safety tuning and access programs. Pricing is starting to track willingness to answer as much as it tracks capability. For anyone building agents, that means evaluating a model now requires looking at its refusal-rate curve and access tier alongside the benchmark scores — that's often what drives a 50% swing in your bill, not the model getting smarter.

References

• OpenAI: GPT-Rosalind-6 System Card Addendum
• OpenAI: Introducing GPT-Rosalind for life sciences research
• OpenAI: Introducing new capabilities to GPT-Rosalind
• OpenAI Help Center: GPT-Rosalind for life sciences research
• OpenAI: API Pricing
• i-scoop: GPT-6 Rosalind shows OpenAI now sells access, not just models
• beri.net: OpenAI Launches GPT-6 Rosalind at $15/$75 per Million Tokens
• OpenAI: GPT-6 Astra: A new generation of intelligence (comparison baseline, released 2026-09-03)