GPT-Rosalind-6: OpenAI's Biology Model, Access and Pricing
GPT-Rosalind-6 is OpenAI's new model for biology, drug discovery and translational medicine, published on October 30, 2026. It uses the same weights as GPT-6 Astra, OpenAI's flagship reasoning model, but answers more dual-use biology questions, and only vetted organizations can use it. On OpenAI's own life-science tests it scores well above GPT-Rosalind-5.5, and the API price is $15 per million input tokens and $75 per million output tokens.
OpenAI now sells one set of weights under two rule sets: a general model with standard refusals, and a version for labs that pass a review, with fewer refusals on biology.
Key takeaways
- GPT-Rosalind-6 uses GPT-6 Astra's weights, according to OpenAI's system card addendum. It differs in its refusal boundary and in who can use it.
- Against GPT-Rosalind-5.5, MedChemBench v7 rose from 28.9% to 49.6% and GeneBench Pro v13 from 16.7% to 37.7% on OpenAI's internal tests.
- OpenAI rates it High for biological risk, below Critical. None of its four Critical-level biology design evaluations reached the threshold.
- Access requires approval. OpenAI reviews each applicant's mission, use case, governance and access controls, and can restrict or revoke access.
- The API price is $15 input and $75 output per million tokens up to 272K context, 50% more than GPT-6 Astra's $10 and $50 (OpenAI pricing).
What is GPT-Rosalind-6?
GPT-Rosalind-6 is the third model in OpenAI's Rosalind line of life-science models. OpenAI describes it as "a frontier reasoning model for biology, drug discovery, and translational medicine" and offers it to customers under the product name GPT-Rosalind-6 Research (system card addendum). In the API it appears as gpt-6-rosalind (pricing page).
The series started in April 2026 with the first GPT-Rosalind research preview. GPT-Rosalind-5.5 followed on June 3, 2026, and was trained further from GPT-5.5 (GPT-Rosalind-5.5 system card).
Version 6 is built differently. OpenAI says it uses the same weights as GPT-6 Astra, with no added biology training data or post-training. A dual-use query is a question with legitimate research value that could also help someone cause harm, such as details of how a pathogen enters cells. On those queries Rosalind refuses less often than Astra. It still refuses requests for biological weaponization and keeps Astra's refusals on other safety topics.
Since the weights are Astra's, most of the improvement over GPT-Rosalind-5.5 comes from the newer base model, not from new biology training. An analysis by i-SCOOP makes the same point.
How it compares with GPT-Rosalind-5.5 and GPT-6 Astra
The table pulls together the figures OpenAI has published. Benchmark numbers are OpenAI's internal results at the highest reasoning setting and have not been independently reproduced.
| GPT-Rosalind-5.5 | GPT-Rosalind-6 | GPT-6 Astra | |
|---|---|---|---|
| Base | Trained further from GPT-5.5 | Same weights as GPT-6 Astra | Flagship model |
| Published | June 3, 2026 | October 30, 2026 | not covered here |
| Access | Trusted access | Trusted access | General API and ChatGPT |
| MedChemBench v7 (99 problems) | 28.9% | 49.6% | not reported |
| GeneBench Pro v13 (129 problems) | 16.7% | 37.7% | not reported |
| LifeSciBench Gold (94 problems) | 44.0% | 61.3% | not reported |
| Biorisk knowledge, raw score | 79.2% | 78.3% | 63.3% |
| API price per 1M tokens, input / output | $5 / $25 (gpt-rosalind-research) | $15 / $75 | $10 / $50 |
Sources: GPT-Rosalind-6 addendum, GPT-Rosalind-5.5 system card, OpenAI API pricing. Prices are for prompts up to 272K tokens.
What the benchmarks measure
- MedChemBench v7 grades answers to medicinal chemistry problems against a rubric and reports the mean rubric score.
- GeneBench Pro v13 measures whether the model solves a genomics task on a single attempt. OpenAI notes the test configuration differs from production in some ways.
- LifeSciBench Gold is a broad life-science set. OpenAI excluded six of its 100 problems, leaving 94.
All three are OpenAI's own benchmarks. They show how much one OpenAI model improved on another and say nothing about how Rosalind ranks against models from Anthropic, Google or other labs.
Why OpenAI rates it High, not Critical
OpenAI's Preparedness Framework sorts risk into capability levels. "High" is the level at which OpenAI requires safeguards before a launch (GPT-Rosalind-5.5 system card). "Critical" is the level above it. OpenAI treats GPT-Rosalind-6 as High in the biological and chemical domain, below Critical (addendum).
On biorisk knowledge, Rosalind scored 78.3% with a 9.4% refusal rate. Adjusted for refusals, it scores 85.0% against GPT-6 Astra's 90.0%, so Rosalind answers more of the questions while Astra is a bit more accurate on the ones it answers.
On TroubleshootingBench, a set of wet-lab troubleshooting questions, it answered 491 of 832 correctly, about 59%. The indicative High threshold is 36.4%.
OpenAI also ran four tests of biological design ability, including AAV capsid packaging and SHP2 protein function prediction. None reached its indicative Critical threshold.
On cybersecurity, the addendum relies on GPT-6 Astra's assessment, which OpenAI rates Critical. Rosalind keeps Astra's standard cyber safeguards, including real-time blocking, and adds a more conservative cyber refusal boundary. In OpenAI's historical cyber jailbreak test, Rosalind succeeded on 14.3% of attempts against 15.7% for Astra, where lower is better.
OpenAI lists its own limits too. Critical thresholds are preliminary estimates. The evaluations do not fully test the ability to cause biological harm. Results cannot be compared directly with earlier system cards because the benchmarks and grading have changed.
Who can get access
GPT-Rosalind-6 is not part of a ChatGPT subscription. Trusted access is OpenAI's vetting program for sensitive models: an organization applies, OpenAI checks who it is and how it will use and control the model, and only approved organizations get in. OpenAI offers it to "qualified scientists, research institutes, companies and government partners" and reviews each applicant before granting access (addendum). The review covers:
- A legitimate life-science mission and a credible use case.
- Research and operational governance.
- The ability to control who inside the organization uses the model.
Non-government applicants go through business verification and compliance screening. Government bodies have a separate route for biosecurity, biosafety, public-health preparedness and biodefense work. Eligibility follows the rules set for GPT-Rosalind-5.5; this launch does not widen them.
Approved customers take on duties of their own. They must limit access to authorized users, apply least-privilege controls, remove access people no longer need, and have a qualified person review outputs before they feed into experiments, publications or other consequential decisions. OpenAI monitors use and can restrict or revoke access.
OpenAI's help center describes two tiers, Discovery and Research. Approval for Discovery does not carry over to Research, and Research access is not available in ChatGPT's regular chat tab. US university labs without an enterprise agreement can apply through the Rosalind Academic Plan, which gives eligible labs up to 30 Business Standard seats at no cost for up to 12 months, still subject to trusted-access approval.
Pricing
On OpenAI's API pricing page, gpt-6-rosalind costs:
| Context length | Input | Cached input | Output |
|---|---|---|---|
| Up to 272K tokens | $15.00 | $1.50 | $75.00 |
| Over 272K tokens | $30.00 | $3.00 | $112.50 |
Prices are per million tokens. For comparison, GPT-6 Astra costs $10 input and $50 output in the same tier, and the older gpt-rosalind-research and gpt-rosalind-discovery models cost $5 and $25. Regional data-residency endpoints add 10%.
A lab therefore pays 50% more than Astra's list price for the same weights, in exchange for the looser refusal boundary on dual-use biology.
What it means for labs and biotech teams
Teams already on GPT-Rosalind-5.5 get large gains on medicinal chemistry and genomics tasks, though these are OpenAI's numbers on OpenAI's tests. Anyone moving production work over should run their own evaluations on their own data first, and budget for output that costs three times as much as gpt-rosalind-research.
Teams not yet approved face the same eligibility rules as before. They need governance, access controls and human review of outputs in place before they apply.
The launch also shows a change in method. GPT-Rosalind-5.5 was trained further from GPT-5.5; version 6 takes the general model as is, relaxes refusals in one domain, and relies on vetting to control who uses it. We saw the same lab tighten its own training setup after a sandbox incident in OpenAI's move to offline frontier training, and its general-purpose line keep moving with GPT-6.1 Terra.
FAQ
Is GPT-Rosalind-6 available in ChatGPT Plus or Pro?
No. GPT-Rosalind-6 is available only to organizations approved through OpenAI's trusted-access program. OpenAI's help center says Research-tier access is not available in ChatGPT's regular chat tab.
Is GPT-Rosalind-6 a separately trained model?
No. OpenAI's system card addendum says GPT-Rosalind-6 uses the same weights as GPT-6 Astra. The differences are a more permissive refusal boundary for dual-use biology questions and restricted access.
How much does GPT-Rosalind-6 cost?
On OpenAI's API pricing page, gpt-6-rosalind costs $15 per million input tokens and $75 per million output tokens for prompts up to 272K tokens, and $30 and $112.50 above that. Cached input costs $1.50 per million tokens in the lower tier.
Is GPT-Rosalind-6 dangerous?
OpenAI rates it High for biological and chemical capability, which requires safeguards, but below Critical. None of its four Critical-level biology design evaluations reached the threshold. It still refuses requests for biological weaponization.
Can individual researchers apply?
Access goes to organizations, not individuals. OpenAI reviews the organization's mission, use case, governance and access controls. US university labs can apply through their principal investigator under the Rosalind Academic Plan.
Sources
- OpenAI Deployment Safety Hub: GPT-Rosalind-6 System Card Addendum
- OpenAI Deployment Safety Hub: GPT-Rosalind-5.5 System Card
- OpenAI API pricing
- OpenAI Help Center: GPT-Rosalind for life sciences research
- OpenAI: Introducing GPT-Rosalind for life sciences research
- OpenAI: Bringing GPT-Rosalind to academic research labs
- i-SCOOP: GPT-6 Rosalind shows OpenAI now sells access, not just models
