GPT-6 Rosalind is OpenAI's October 2026 life sciences model for approved Research organizations, scoring 61.3% on LifeSciBench Gold where GPT-Rosalind-5.5 scored 44.0%. It suits bioinformaticians and early drug discovery teams that already hold Rosalind Research approval, and it replaces the 5.5 model only where the newer biology boundary breaks no existing workflow.
Can you use it?
- ChatGPT Work and Codex (GPT-6 Rosalind Research): gated (Approved Rosalind Research organizations; Enterprise workspaces start in a two-week Admin Preview with the model off by default). Not in the Chat tab; Discovery approval does not grant Research access.
- OpenAI API (gpt-6-rosalind): gated (Separately approved API organizations, internal research use only). Not for customer-facing products or external commercial applications.
- ChatGPT Chat tab: no. OpenAI states GPT-6 Rosalind Research is not available in the Chat tab.
- Free plan: not published.
- Gate: OpenAI trusted-access program (Rosalind Research), Organization applies through the GPT-Rosalind access request and is verified; an owner or admin accepts terms and activates. Discovery approval does not grant Research, and API access is approved separately..
Checked 3 Nov against OpenAI's own pages. Compare access across frontier models
What changed
GPT-6 Rosalind is OpenAI's life sciences reasoning model, released in October 2026 on the same weights as GPT-6 Astra. OpenAI reports 37.7% on its GeneBench Pro v13 genomics evaluation against 16.7% for GPT-Rosalind-5.5, with fewer refusals on dual-use biology. Only organizations approved through OpenAI's trusted-access program can use it.
Where it sits
- $30.00/M$ per 1M tokensBlended price (3:1)Lower is better#91 / 92peer median $1.70/Mvendor price, checked by Hokai
Pricier than 99% of the 92 GA models with a published price, and one of 36 that document a zero-data-retention option.
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Provider: OpenAI · Family: GPT-Rosalind
Input modalities: text, tool-calls · Output: text, tool-calls
About GPT-6 Rosalind
OpenAI released GPT-6 Rosalind on October 30, 2026, after the original GPT-Rosalind on April 16, 2026 and the GPT-Rosalind-5.5 update on June 3, 2026. OpenAI's release notes call the product GPT-6 Rosalind Research, its system card calls it GPT-Rosalind-6, and the API identifier is gpt-6-rosalind. It is not a separate training run: the system card states that it uses the same weights as GPT-6 Astra, that OpenAI added no training data, and that it did no additional post-training. What changes is the boundary around those weights, with fewer refusals on dual-use biology questions in exchange for access limited to approved organizations.
OpenAI's system card compares it with GPT-Rosalind-5.5 on three life sciences benchmarks, re-running the older model on the same problems, so the 5.5 figures differ from those published at its June launch. At the highest reasoning budget tested, the mean rubric score on MedChemBench v7 (99 medicinal chemistry problems) was 49.6% against 28.9%. Single-attempt success on GeneBench Pro v13 (129 genomics and quantitative biology problems, run in a benchmark-specific configuration that differs from production) was 37.7% against 16.7%. LifeSciBench Gold, 94 retained tasks contributed by practicing life scientists, averaged 61.3% against 44.0%. These are OpenAI's own results on benchmarks it built, and Hokai found no independent evaluation as of November 3, 2026, so the benchmark table on this page stays empty instead of borrowing scores from the parent model.
Under its Preparedness Framework, OpenAI treats the launch as High capability in Biological and Chemical, below the Critical threshold, and as Critical in Cybersecurity, a rating it carries over from Astra. Adjusted for refusals, the model passes the indicative High thresholds on three of its four High-capability biology evaluations, and none of the four Critical-capability evaluations met its threshold. Refusals moved as designed against the general model: on the Biorisk knowledge set it refused 9.4% of questions, against 29.2% for GPT-6 Astra and 1.8% for GPT-Rosalind-5.5, and its refusal-adjusted consensus score was 85.0% against an indicative High threshold of 80%.
The cyber boundary runs the other way. The model keeps Astra's real-time cyber blocking and adds a more conservative refusal line for malware deployment, credential theft and exfiltration. In a replay of nine known jailbreak attacks across ten ExploitBench tasks with real-time blocking switched off, 14.3% of tasks succeeded against Rosalind and 15.7% against Astra, counting seven attacks because two ended in errors or turn limits. Authorized exploit work is better served by GPT-5.6-Cyber, which OpenAI gates through its separate Daybreak program.
Access is the real product. GPT-6 Rosalind runs in ChatGPT Work and Codex and, for separately approved organizations, in the API for internal research tools and workflows. It is not in the Chat tab, not open to customer-facing products and not on Amazon Bedrock, and Rosalind Discovery approval does not carry over to Research. Inside Enterprise workspaces the first two weeks run as an Admin Preview, with the model switched off until an admin turns it on. GPT-5.5 Rosalind Research stays selectable, and Rosalind Discovery remains on the 5.5 model.
OpenAI publishes a price for gpt-6-rosalind but no model page, so Hokai lists no context window, output limit, rate limits or modality list for it. The GPT-6 Astra page lists a 1,050,000-token window and 128,000 output tokens, but OpenAI does not say those carry over to this deployment. Per-token pricing is three times that of the GPT-Rosalind-5.5 models, which build on GPT-5.5 and stay available for work that does not need the newer weights. OpenAI's standard pricing page also lists GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, the other GPT-6 models for general work outside the Rosalind line.
OpenAI describes the Rosalind family as best suited to bioinformaticians, computational biologists and early discovery biologists, and as most valuable today for early discovery and hypothesis generation: target biology, mechanism work, literature synthesis and omics interpretation. Approved customers must apply qualified human review before outputs inform experiments, publications or consequential decisions. To see where else OpenAI's flagship weights are open, browse OpenAI's models, the access tracker and the other proprietary models.
Pricing
USD per 1M tokens as listed on OpenAI's pricing page on November 3, 2026: 15 in, 75 out and 1.50 cached at the short-context rate. The Long-context rate is exactly twice the input price and 1.5 times the output price. OpenAI's Rosalind row does not say where Long context begins; the [GPT-6 Astra](/hub/models/gpt-6-astra) page sets that line at 272K input tokens for Astra. Billing follows approval, and API use is limited to approved internal research.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.450 | $0.075 | $0.525 |
| Support reply | $0.030 | $0.022 | $0.052 |
| One coding agent run | $3.00 | $1.50 | $4.50 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Same weights as GPT-6 Astra: The system card states that OpenAI added no training data and did no extra post-training, so every difference from Astra comes from the access boundary and safeguards.
- Looser dual-use biology boundary: Answers some dual-use biology questions that Astra may refuse in general access, while still refusing requests that would meaningfully enable biological weaponization.
- Work, Codex and API surfaces: Runs in ChatGPT Work, in Codex alongside the Life Sciences Research plugin, and in the API for separately approved organizations; it is absent from the Chat tab.
- Two-week Enterprise Admin Preview: Enterprise workspaces start with the model off by default for two weeks, and an owner assigns a custom role that turns it on for approved users.
- Stricter cyber refusal line: Adds a tighter refusal boundary for malware deployment, credential theft and exfiltration beside Astra's live blocking layer, then tests it with replayed jailbreak attacks.
Pros
- Largest published gains over its predecessor are in genomics and medicinal chemistry, the two drug-discovery areas OpenAI tuned the Rosalind line for.
- Refuses fewer dual-use biology questions than the general model on the same weights, which is the practical reason a research team asks for it.
- Access, retention and admin controls are documented in detail, including role-based enablement and separate retention approval.
Cons
- Gated: needs Rosalind Research approval for the organization and the user, and Discovery approval does not carry over.
- Costs three times the earlier Rosalind models per token, and long prompts bill higher again.
- No spec sheet: context size, output cap, throughput limits and supported input types are all unpublished.
- Every benchmark is OpenAI's own, on benchmarks it built, with no independent evaluation found.
- API use is limited to approved internal research, so you cannot build a customer-facing product on it.
Frequently Asked Questions
How much do you pay for GPT-6 Rosalind?
At the standard rate, OpenAI charges $15.00 per 1M input tokens and $75.00 per 1M output tokens, with cached input at $1.50. Long-context requests move to $30.00 input and $112.50 output, and cache writes are listed at $18.75 for 30 minutes and $37.50 for 12 hours. Those rates are three times the $5.00 and $25.00 OpenAI charges for the earlier GPT-Rosalind models, and there is no free tier.
Is GPT-6 Rosalind better than GPT-Rosalind-5.5?
On OpenAI's like-for-like re-run it wins all three life sciences benchmarks it published: about 1.7 times the medicinal chemistry score, 2.3 times the genomics success rate and 1.4 times the applied research score. The price is a tighter boundary, because Biorisk knowledge refusals rose from 1.8% to 9.4% and the unadjusted consensus score slipped from 79.2% to 78.3%. Move to GPT-6 Rosalind for genomics and medicinal chemistry work, and keep an existing workflow on 5.5 if the new refusals break it.
Who can use GPT-6 Rosalind and how do you get access?
Organizations with Enterprise or Business agreements apply through OpenAI's GPT-Rosalind access request, OpenAI verifies the organization, and a verified owner or admin then accepts terms and activates it, so approval and activation are separate steps. Select government entities have a separate pathway for approved biosecurity, public-health preparedness and biodefense work, and U.S. university labs can apply through the Rosalind Academic Plan, which carries no API access. A ChatGPT subscription or workspace seat grants nothing on its own.
Does GPT-6 Rosalind train on the data you send it?
OpenAI says it does not train on enterprise customer data by default, and that the protections you get depend on your approved product, agreement and workspace configuration. Model access and data retention are approved separately, so Rosalind approval does not by itself turn on Enhanced Privacy or Zero Data Retention. The Business-based Rosalind Academic Plan does not automatically inherit enterprise privacy terms, so academic labs should read its own data-handling page before uploading research data.
Who should use GPT-6 Rosalind, and who should pick something else?
Choose it if your organization already holds Research approval and the work is early discovery: target biology, genomics and omics interpretation, medicinal chemistry or wet-lab troubleshooting, with qualified scientists reviewing outputs. Without approval, GPT-6 Astra is the generally available route on the same weights at $10.00 and $50.00 per 1M tokens, though with tighter biology refusals. Everyday literature review and protocol reading fit Rosalind Discovery on the 5.5 model at $5.00 and $25.00, and customer-facing products are outside the API terms.
Top Alternatives
- GPT-6 Astra: Pick GPT-6 Astra if your organization has no Rosalind Research approval, since it runs on the same weights through the generally available API.
- GPT-5.6-Cyber: Pick GPT-5.6-Cyber for authorized exploit work, because GPT-6 Rosalind adds a more conservative cyber refusal line.
- GPT-5.5: Pick GPT-5.5 for general API work, since GPT-6 Rosalind is limited to approved internal life sciences research.