GPT-6 Rosalind charges life-sciences teams 1.5 times Astra's token price for the same weights, and the premium buys fewer refusals on dual-use biology (cyber refusals get stricter), only for organizations accepting OpenAI's control obligations. OpenAI's system card addendum, published October 30, 2026, is direct about the model: "We did not add training data or perform additional post-training for this release."
So judge it on three things: how often GPT-6 Astra refuses your actual research prompts, whether your organization can meet the review and access duties, and whether that refusal cost justifies the price. The benchmark gains over GPT-Rosalind-5.5 are real in OpenAI's own testing, but they come with the Astra base model.
What OpenAI Shipped on October 30
GPT-6 Rosalind is GPT-6 Astra's weights released under a looser biology refusal boundary and a trusted-access program. The API changelog describes it as "specialized biological reasoning with scientific tool use," and the addendum calls it "a frontier reasoning model for biology, drug discovery, and translational medicine." The names vary by document: the system card says GPT-Rosalind-6, the commercial offer is GPT-Rosalind-6 Research, and the API identifier is gpt-6-rosalind.
Trusted access is OpenAI's approval process for restricted models. Per the addendum, OpenAI reviews an applicant's life sciences mission, use case, governance and ability to control access. Non-government applicants also go through business verification and compliance screening, and select government entities have a separate route for approved biosecurity, biosafety, public-health and biodefense work.
The release adds no new eligibility. "This release does not broaden eligibility beyond prior announcements, such as the Rosalind Academic Plan," the addendum says. That plan gives approved US university labs up to 30 free ChatGPT seats for up to 12 months, and the model itself still needs separate approval. i-SCOOP reports that a GPT-Rosalind Discovery approval does not cover the new model, and that approved users reach it through Codex, ChatGPT's work environment or the API.
We covered the original GPT-Rosalind in April, when its pitch rested on expert comparisons. Its predecessor here, GPT-Rosalind-5.5, was built on GPT-5.5.
GPT-6 Rosalind Costs $15 Input and $75 Output per Million Tokens
OpenAI's price is now published, and it settles a question launch coverage left open. i-SCOOP and Data Studios both reported no gpt-6-rosalind row on the pricing page as of October 31, while FourWeekMBA read $15 and $75 from it the same day. When checked on November 1, 2026, OpenAI's pricing page listed the row.
| Model (standard, per 1M tokens, up to 272K input) | Input | Cached input | Output |
|---|---|---|---|
gpt-6-rosalind |
$15.00 | $1.50 | $75.00 |
gpt-6-astra |
$10.00 | $1.00 | $50.00 |
gpt-rosalind-research |
$5.00 | $0.50 | $25.00 |
Above 272K input tokens the new model's rates rise to $30 input, $3 cached and $112.50 output, against $20, $2 and $75 for Astra. Regional processing carries a 10% uplift on eligible models, which would make it $16.50 and $82.50, though i-SCOOP found nothing published on whether this model qualifies for EU regional processing.
Take an illustrative research agent run of 200K uncached input tokens and 20K output. That costs $1.50 on gpt-rosalind-research, $3.00 on Astra and $4.50 on the new model. At 10,000 runs a month you are comparing $15,000, $30,000 and $45,000. Moving from gpt-rosalind-research triples the bill before any caching.
OpenAI's legal configuration carries no such premium for now. FourWeekMBA lists gpt-6-astra-law at the standard $10 and $50, and OpenAI's pricing page calls that promotional, available at least through December 31, 2026. We covered Astra for Law's benchmark claims in September.
Fewer Refusals Are What You Are Buying
On OpenAI's benchmarks the refusal gap looks large enough to matter on legitimate work, and the safety margin on sensitive prompts falls by more than twice as many points. Data Studios tabulates the system card's figures this way:
| Biosecurity test | GPT-6 Rosalind | GPT-6 Astra |
|---|---|---|
| Benign biology, fully answered | 100% | 91.6% |
| Dual-use, passes safety grader | 84.6% | 97.8% |
| Sensitive, passes safety grader | 79.0% | 99.8% |
The strongest case for paying is in the first row. If Astra declines or caveats about one benign biology question in twelve, an unattended agent can stall partway through a literature or protocol workflow, and the scientist who has to step in can cost more than the token premium. Teams seeing that pattern have a straightforward reason to apply.
The other rows are the price of that fix. On sensitive material about one Rosalind response in five fails the safety grader, against almost none of Astra's (OpenAI notes a flagged response is not necessarily unsafe), and OpenAI's answer is the access program. The addendum says these grader scores come from prompts related to biological weaponization, and that for this deployment it mainly considered a benchmark called Biosecurity - Severe. Data Studios' table shows no biology row under that name, so ask your account team for that score. Your controls absorb part of what the model no longer stops. It still refuses requests for biological weaponization, and the addendum says it keeps refusals on other safety topics.
Capability moved far less than willingness to answer. On OpenAI's biorisk knowledge test, Rosalind beat Astra 78.3% to 63.3% before adjustment, with refusal rates of 9.4% and 29.2%. After adjusting for refusals the order reverses, 85.0% to 90.0%. i-SCOOP puts it plainly: the advantage "comes from answering more often, not from answering better." Shared weights would predict that result.
GPT-Rosalind-5.5 Customers Get Better Scores and More Refusals
Moving up from GPT-Rosalind-5.5 improves all three of OpenAI's published research benchmarks and raises the refusal rate on the biorisk test. The addendum's figures, at the highest reasoning budget:
| Benchmark | GPT-6 Rosalind | GPT-Rosalind-5.5 |
|---|---|---|
| MedChemBench v7 (99 problems) | 49.6% | 28.9% |
| GeneBench Pro v13 (129 problems) | 37.7% | 16.7% |
| LifeSciBench Gold (94 problems) | 61.3% | 44.0% |
Read those as vendor claims. OpenAI ran the tests and reported the scores, and GeneBench used "a benchmark-specific configuration that differs from production." The 5.5 scores do use the same benchmark versions and problem sets as the new model. i-SCOOP notes they have not been independently reproduced and that the gains come from the Astra base. Even the best score here, on a set of 94 problems, is 61.3%, which is why the human-review duty below has practical weight.
Two numbers cut the other way. On the biorisk knowledge test, GPT-Rosalind-5.5 refused 1.8% of questions and scored 79.2% against the new model's 9.4% and 78.3%. Data Studios also reports 5.5 did better on phage-plasmid co-evolution, with a negative log-likelihood of 11.81 against 13.41, where lower is better. A workflow tuned on 5.5 should be re-tested before you switch it, with gpt-rosalind-research still on the price list at a third of the cost, though the page does not say which version that identifier serves.
The Access Terms Move the Control Burden to You
Approval comes with duties that reach into your research process, and OpenAI keeps a broad right to take the model away. Under the addendum's customer section, approved enterprises must limit access to authorized users with a legitimate need, apply least-privilege controls, onboarding and insider risk controls, and remove access when the need ends.
The widest duty is human review: "Customers must apply qualified human review before model outputs inform experiments, publications, or consequential decisions." The addendum does not define "qualified" or say what a reviewer must document. An agent that proposes a compound and queues it for a lab run has an output informing an experiment, so a closed-loop design needs a named review step before anything reaches the instrument. If you already run validated systems, the context-of-use work for GxP is the closest template, and Enveda's batch record reviewer shows one way to keep a quality unit in the loop.
Revocation has no stated process. "OpenAI can restrict or revoke access for misuse" and a mention of monitoring and enforcement processes are all the addendum says, with no triggers, notice period or appeal, and AI Weekly flagged the same open question. A pipeline with no fallback model has a single point of failure.
Use is also limited. The pricing page restricts access to "approved internal research through the trusted-access program," so a CRO or software vendor planning a customer-facing product should get a written answer first. And red-teaming is unfinished: OpenAI says it is repeating its biosecurity exercise "now that it is deployed" and will not disclose details, which your security team will have to accept or raise directly.
The security profile needs its own review. OpenAI rates the model High for biological and chemical capability, below Critical, and Critical for cybersecurity, with real-time blocking and a more conservative cyber refusal boundary. One agent test moved the wrong way: unintended engagement with external agent messages came in at 2.5% against 0.0% for Astra, in a table the addendum calls a descriptive comparison because versions, settings and grading differ across some runs. OpenAI gives only a one-line definition of the test, so treat it as a question for your account team if your agents read outside content.
What to Do Before You Apply
This Week:
- Pull 200 real prompts from your research teams and run them on GPT-6 Astra with your research lead scoring each one as answered, caveated or refused. OpenAI's model page lists no approval step for Astra, only usage-tier rate limits, and this gives you the only refusal rate that applies to your work.
- Ask your OpenAI account team whether your current Rosalind approval covers
gpt-6-rosalind, and get the answer in writing.
This Month:
- Price the gap. Multiply refused prompts by the scientist time each one costs, then compare it with your own token volume at $15 and $75 against $10 and $50.
- Have your research informatics lead draft the review step: who qualifies as a reviewer, what they sign, and where the record lives. Ask your security lead to map least-privilege roles and the offboarding path to your identity provider.
- Re-run your GPT-Rosalind-5.5 evaluation set on any candidate replacement before changing a production workflow, and record refusals separately from wrong answers.
Before You Sign:
- Ask OpenAI for the revocation triggers, notice terms and data handling for this model, including EU regional processing if your data requires it. Have counsel review the answers alongside any genetic or health data in scope.
- Name a fallback model for every workflow and test it once. Anthropic sells into the same market with Claude Science, so a second vendor is a realistic option to evaluate.
The Bottom Line
OpenAI now prices its Astra configurations by what surrounds the model. The same pricing page sells Astra at six times the standard rate for Ultrafast and 1.5 times for Rosalind's refusal boundary and access program. OpenAI's purchase of Orchestra Bio suggests life sciences will keep getting its own products, so expect this kind of decision to recur. Buy on your measured refusal rate and your readiness to meet the review duty. Start with the 200-prompt test on Astra, and take its result to your account team.