OpenAI prices life-sciences GPT-Rosalind-6 at 1.5x GPT-6 Astra

OpenAI has released GPT-Rosalind-6, a life-sciences version of GPT-6 Astra that runs on the same weights. A new gpt-6-rosalind row on OpenAI’s API pricing page lists it at $15.00 per million input tokens and $75.00 per million output tokens. GPT-6 Astra is listed at $10.00 and $50.00, so the same weights cost 1.5 times as much for both input and output.

OpenAI’s system card addendum, published on 30 October, says no new training went into the release. Customers who pay the premium get different access terms and refusal settings, not a retrained model.

Same weights as GPT-6 Astra, with a different policy

The addendum is explicit about what did not change:

“We did not add training data or perform additional post-training for this release.”

FourWeekMBA cites that line from the addendum. Post-training is the stage after initial training in which a model is tuned for behaviour, such as following instructions or refusing certain requests.

The changes are in who may use the model and what it will answer. OpenAI writes that “Rosalind has fewer refusals on dual-use biological queries but is only made available via trusted access”. Dual-use queries are questions whose answers could serve legitimate research or cause harm.

The addendum also describes a trade-off in the other direction. GPT-Rosalind-6 applies “a more conservative cyber refusal boundary” than GPT-6 Astra, its base model. So the variant is more permissive than Astra on biology and stricter on cybersecurity.

Model (pricing page row)	Input, per million tokens	Output, per million tokens	Multiple of GPT-6 Astra
GPT-6 Astra (gpt-6-astra)	$10.00	$50.00	1.0x
GPT-Rosalind-6 (gpt-6-rosalind)	$15.00	$75.00	1.5x
Legal configuration (gpt-6-astra-law)	$10.00	$50.00	1.0x
gpt-rosalind-research / gpt-rosalind-discovery	$5.00	$25.00	n/a

All prices are as FourWeekMBA read them from OpenAI’s API pricing page on 31 October. The multiples are FourWeekMBA’s arithmetic.

Two premiums, two choices on the pricing page

The same pricing page lists gpt-6-astra-law, described as a legal model configuration, at $10.00 input and $50.00 output. That is the same price as GPT-6 Astra. OpenAI is therefore charging a premium for the life-sciences configuration of its flagship weights but not for the legal one.

FourWeekMBA’s reading is that “OpenAI is pricing the access terms and refusal settings, since the model underneath is the one GPT-6 Astra customers already buy”. That is the outlet’s interpretation; neither OpenAI document quoted in the report explains the premium.

There is also a naming gap. According to FourWeekMBA, the pricing page does not use the system card’s name, GPT-Rosalind-6. The outlet treats the gpt-6-rosalind row as GPT-Rosalind-6 for two reasons:

• When it read the same Rosalind table on 11 October, the table held only gpt-rosalind-research and gpt-rosalind-discovery.
• The new row follows the gpt-6-astra naming pattern.

Builders should treat the link between the row and the model as a close reading of OpenAI’s pages, not as an explicit statement from OpenAI.

The two older entries are priced at $5.00 and $25.00 and have been billed since 5 October. TechFlow reported on 10 October that gpt-rosalind-discovery had appeared on the pricing page before any public release. Its price matched gpt-rosalind-research, and its purpose was unclear; some speculated it was meant for drug discovery.

Who can get GPT-Rosalind-6, and on what terms

The addendum says GPT-Rosalind-6 “is available only to approved customers”. Before granting access, OpenAI “reviews whether the applicant has a legitimate life sciences or related mission, a credible use case, appropriate research and operational governance, and the ability to safely control access to the model”.

The routes to access are:

• Non-government applicants: they also go through business verification and compliance screening.
• Government applicants: a separate pathway covers “select government entities working on approved biosecurity, biosafety, public-health preparedness, and biodefense use cases”.

OpenAI says the release “does not broaden eligibility beyond prior announcements, such as the Rosalind Academic Plan”.

The pricing page describes access in narrower terms. It says access to the Rosalind models “is limited to approved internal research through the trusted-access program”. The addendum’s “approved customers” wording and the pricing page’s “approved internal research” wording are not identical. Teams should confirm with OpenAI what counts as permitted use before planning a product around the model.

Approved customers also take on obligations. They must limit access to authorised users with a legitimate need. They must also “apply qualified human review before model outputs inform experiments, publications, or consequential decisions”, the addendum says.

The sources do not say which surfaces carry GPT-Rosalind-6. The earlier GPT-Rosalind model was offered through ChatGPT, Codex and the API. For this release, only the API pricing row is documented in the coverage.

The safeguards OpenAI describes, and what it does not publish

Refusals. OpenAI says the model still refuses “requests for biological weaponization” and keeps refusals on other safety topics.

Risk ratings. OpenAI’s Preparedness Framework is its internal system for rating model risk. Under it, OpenAI treats the launch as High capability in the biological and chemical domain, which is below the Critical level. It treats GPT-Rosalind-6 as Critical in cybersecurity. That rating relies on GPT-6 Astra’s assessment, not a separate one for the variant.

Jailbreak resistance. The one comparative number in the addendum comes from OpenAI’s own testing. OpenAI replayed known jailbreak attacks with real-time blocking switched off. A jailbreak is a prompt crafted to get around a model’s safety rules. GPT-Rosalind-6 had a 14.3% task success rate against 15.7% for GPT-6 Astra; lower is better. This is an internal evaluation, and no independent replication appears in the sources.

Enforcement. OpenAI says it “can restrict or revoke access for misuse”. It says it tested its monitoring and enforcement through biosecurity red-teaming on GPT-Rosalind-5.5 and is repeating that work for GPT-Rosalind-6. Red-teaming is structured adversarial testing. OpenAI says it is not publishing the details.

Outside readers therefore have the ratings and one jailbreak figure, but not the red-team findings behind OpenAI’s biosecurity claims.

How GPT-Rosalind got here

OpenAI’s news feed dates the first GPT-Rosalind release to 16 April 2026. At launch, R&D World described it as OpenAI’s first life-sciences model, named for the chemist Rosalind Franklin. It was offered as a limited research preview in ChatGPT, Codex and the API for qualified customers under OpenAI’s trusted-access program.

OpenAI framed that model around evidence synthesis, hypothesis generation, experimental planning and multi-step research tasks. It called it a “frontier reasoning model” and the first release in a “GPT-Rosalind life sciences model series”. Early access partners named at the time included Moderna, Retro Biosciences, Genentech and Thermo Fisher Scientific. Scalevise’s October summary of OpenAI’s materials also lists Novo Nordisk, Amgen and the Allen Institute among the life-sciences organisations OpenAI works with. It describes the model as sitting within Rosalind Workbench, a workspace for examining sequences, synthesising evidence, comparing targets and planning experiments.

According to FourWeekMBA, the series then grew as follows:

• May: a biodefense programme.
• June: new capabilities.
• September: the research preview ended.
• October: the model was listed on the API pricing page.
• 15 October: an academic plan.

By the outlet’s count, GPT-Rosalind-6 arrives 197 days after the first release.

Scalevise notes that the access programmes are easy to confuse. The ChatGPT for Academic Researchers offer gives eligible faculty and postdocs a workspace of up to five seats for 12 months at no subscription cost, with eligibility verified through SheerID. It does not by itself confirm GPT-Rosalind access. Rosalind Biodefense offers sponsored GPT-Rosalind access and launch support for qualifying biodefense and pandemic-preparedness work.

What it means for builders

For most API developers, GPT-Rosalind-6 is not a model they can sign up for. Access is gated by mission review, business verification and governance checks. A team that qualifies gets two things over GPT-6 Astra:

• more willingness to answer dual-use biology questions;
• a stricter stance on cybersecurity prompts.

Because the weights are the same, teams should not expect better scientific reasoning than Astra. The addendum rules out new training for this release.

That shapes the build-or-buy question. If an Astra-based pipeline already handles a team’s literature or data work and does not hit biology refusals, the 1.5x price buys mainly policy, not capability. If refusals on legitimate dual-use research are the blocker, the premium pays for a model configured to answer those questions inside a vetted programme.

The human-review requirement also matters for product design. Any workflow in which model output flows directly into experiments, publications or consequential decisions would need a qualified reviewer in the loop to meet OpenAI’s terms.

Science models also have uneven strengths. R&D World, testing the plugin layer released with the first GPT-Rosalind, cited Stanford’s 2026 AI Index. That report describes generative AI models as “jagged”: on the ChemBench chemistry evaluation, the best frontier models outperform the best human chemists, yet “they struggle with basic tasks”. The same report says autonomous scientific discovery “remains at an early stage”.

What to watch next

• The model name: whether OpenAI’s pricing page starts using the name GPT-Rosalind-6, confirming FourWeekMBA’s reading of the gpt-6-rosalind row.
• Red-team results: whether OpenAI shares any results from the GPT-Rosalind-6 biosecurity red-teaming it says it is repeating but not publishing.
• Surfaces: whether the new model reaches ChatGPT and Codex, as the earlier GPT-Rosalind did.
• Eligibility: any change to who qualifies. OpenAI says this release does not widen eligibility.
• Other vertical pricing: how OpenAI prices other configurations of GPT-6 Astra. So far, the legal configuration carries no premium and the life-sciences one does.