OpenAI's New Biology Model Is GPT-6 Astra With Fewer Refusals. It Costs 50% More.
GPT-Rosalind-6's system card addendum, published October 30, says it uses the same weights as GPT-6 Astra with no added training data and no additional post-training. It is sold as GPT-Rosalind-6 Research under trusted access to qualified scientists, institutes, companies, and government partners. OpenAI prices gpt-6-rosalind at $15 in, $75 out, and $1.50 cached per million tokens short-context, and $30, $112.50, and $3 long-context, against Astra's $10, $50, $1 and $20, $75, $2. The premium buys a lower refusal boundary on biology, 9.4% against Astra's 29.2% on OpenAI's own table, and an access program, not new weights. We read the system card addendum and the pricing page.
By T.W. Ghost
OpenAI's system card addendum for GPT-Rosalind-6, published today, October 30, 2026, contains the sentence that defines the product: "GPT-Rosalind-6 uses the same weights as GPT-6 Astra. We did not add training data or perform additional post-training for this release." Rosalind-6 is offered as GPT-Rosalind-6 Research, under trusted access, to qualified scientists, research institutes, companies, and government partners after a review of use case, security, and governance.
The pricing page lists gpt-6-rosalind under life sciences models at $15 per million input tokens, $75 output, and $1.50 cached for short context, and $30, $112.50, and $3 for long context. GPT-6 Astra is $10, $50, and $1, and $20, $75, and $2. Same weights, a roughly 50% premium on the short-context rates. What the premium buys is on the addendum's own table: a biorisk-knowledge refusal rate of 9.4% for Rosalind-6 against 29.2% for Astra, and an access gate that decides who gets the lower boundary. It is the second Astra configuration in two weeks to appear on the pricing page under a new name, after Astra for Law. We read the addendum, the pricing page, and the help article. Here is what held up.
What You Are Paying For
| GPT-Rosalind-6 | GPT-6 Astra | |
|---|---|---|
| Weights | Same as Astra; no added training, no additional post-training (OpenAI) | |
| Access | Trusted access only; use case, security, governance review | General |
| Short context, per 1M tokens | $15 in / $75 out / $1.50 cached | $10 / $50 / $1 |
| Long context, per 1M tokens | $30 / $112.50 / $3 | $20 / $75 / $2 |
| Preparedness | High biological and chemical; Critical cybersecurity (inherited); below High for AI self-improvement | Critical cybersecurity |
| Biorisk-knowledge refusal rate | 9.4% | 29.2% |
A biology SKU is not a biology model. The addendum is unusually direct: no new weights. Any capability difference on a life-sciences task is a difference in what the model is allowed to answer and who is allowed to ask.
The refusal rate is the product. Dropping from 29.2% to 9.4% on biorisk knowledge is the point of trusted access, and it is why the access gate exists. The Critical cyber rating comes along because the weights do.
Fifty percent is the price of the program. Review, governance, and a lower boundary cost $5 more per million input tokens and $25 more per million output on short context. Whether that is worth it depends on how many refusals you were hitting on Astra.
The Benchmarks, Sorted by Who Reported Them
| Claim (vs GPT-Rosalind-5.5, highest reasoning budget) | Number | Who |
|---|---|---|
| MedChemBench v7 | 49.6% vs 28.9% | Vendor (OpenAI) |
| GeneBench Pro v13 | 37.7% vs 16.7% | Vendor (OpenAI) |
| LifeSciBench Gold | 61.3% vs 44.0% | Vendor (OpenAI) |
| Biorisk-knowledge refusal rate | 9.4% vs Astra 29.2% | Vendor (OpenAI, system card table) |
The comparison OpenAI prints is against Rosalind-5.5, the previous biology configuration, not against Astra. Since the weights are Astra's, the life-sciences gains over Rosalind-5.5 are GPT-6's gains over GPT-5.5 with a different label on the chart, and all three benchmarks are OpenAI's own.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Same weights as GPT-6 Astra; no added training or post-training | Verified (system card addendum, quoted) |
| Trusted access only; review of use case, security, governance | Verified (addendum) |
| High bio and chem; Critical cyber; below High self-improvement | Verified as OpenAI's Preparedness labels |
| MedChemBench, GeneBench Pro, LifeSciBench Gold scores | Vendor |
| Refusal rate 9.4% vs 29.2% | Vendor (OpenAI table) |
| $15 / $75 / $1.50 short; $30 / $112.50 / $3 long | Verified (OpenAI pricing page) |
| Astra $10 / $50 / $1 short; $20 / $75 / $2 long | Verified (pricing page) |
| Number of organizations approved | Unverified |
Who Should Use It
Apply for Rosalind-6 if your work is hitting Astra's biology refusals and you can pass the review; the lower boundary is real and it is the only thing you are buying.
Skip it if you expected a better model; it is the same model, and the long-context rate is $112.50 per million output tokens. Score it as Astra in any capability comparison, with the system-card caveat attached.
The frontier is elsewhere: on the gates. Anthropic put its biology-adjacent and cyber work behind three verification tiers this month, and OpenAI put its behind a price. Our model comparison has the current standings side by side.
Sources
- OpenAI deployment safety: GPT-Rosalind-6 system card addendum
- OpenAI API pricing
- OpenAI help: GPT-Rosalind for life sciences research
*Not sure whether a trusted-access configuration or the base model fits your lab's budget? Take the free 2-minute quiz and get matched. Then read about the other Astra configuration priced this week: gpt-6-astra-law and the FedRAMP uplift.*