OpenAI releases GPT-Rosalind-6 biology model to vetted labs

Overview

Updated 40 minutes ago

OpenAI began giving approved research groups access to GPT-Rosalind-6 on October 31. The model runs on the same weights as OpenAI's flagship GPT-6 Astra, with no biology-specific training added for this release.

OpenAI's system card rates the model High for biological and chemical capability, below the Critical threshold, and Critical for cybersecurity. It refuses fewer dual-use biology requests than the general-purpose model, and access is limited to organizations that pass OpenAI's review.

Why it matters

A model that can help design experiments and drugs now sits in vetted labs, and its guardrails depend on who gets in.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

$15

Input price per million tokens

API rate for approved Research customers.

$75

Output price per million tokens

API rate for approved Research customers.

49.6%

MedChemBench v7 mean score

GPT-Rosalind-5.5 scored 28.9% on the same 99 problems.

37.7%

GeneBench Pro v13 single-attempt success

GPT-Rosalind-5.5 scored 16.7% on the same 129 problems.

78.3%

Biorisk knowledge consensus score

Below the 80% indicative High capability threshold; GPT-6 Astra scored 63.3%.

9.4%

Refusal rate on biorisk questions

GPT-6 Astra refused 29.2% of the same questions.

Voices

Curated perspectives — historical figures and your fellow readers.

People Involved

Sam Altman

Chief executive, OpenAI

Leading OpenAI's trusted-access strategy

Organizations Involved

OpenAI

Artificial intelligence company

Releasing and governing GPT-Rosalind-6

OpenAI develops the GPT family and runs the trusted-access program that governs Rosalind.

Timeline

September 2026 October 2026

4 events Latest: Today

Tap a bar to jump to that date

October 2026
OpenAI releases GPT-Rosalind-6
Today Product launch

Approved Research organizations and separately approved API customers gain access at $15 per million input tokens and $75 per million output tokens.
Today
System card addendum published
Safety evaluation

OpenAI rates GPT-Rosalind-6 High for biological and chemical capability, below Critical, and Critical for cybersecurity.
Yesterday
API changelog lists GPT-6 Rosalind
Product availability

OpenAI's API changelog adds the model for approved life-sciences research customers.
Yesterday
September 2026
GPT-Rosalind exits research preview
Product launch

OpenAI opens GPT-Rosalind to eligible organizations worldwide through its trusted-access program.
September 11th, 2026

Scenarios

Trusted access expands without a documented misuse case

Likely Resolves by Apr 30, 2027

Discussed by: OpenAI's deployment safety documentation and life-sciences access pages

OpenAI approves more Research organizations and keeps the model behind business verification, security review, and human oversight. No confirmed case of an approved customer using the model for prohibited biological work appears in public records. The model becomes a standard tool for vetted drug-discovery and genomics teams.

A documented misuse case forces tighter controls

Possible Resolves by Oct 31, 2027

Discussed by: Biosecurity researchers and OpenAI's own safeguards section

An approved customer uses GPT-Rosalind-6 for biological work outside its stated scope, and the case becomes public through an OpenAI disclosure or a government investigation. OpenAI responds by suspending the customer, adding classifier checks, or restricting the model's dual-use answers.

OpenAI opens GPT-Rosalind-6 to general ChatGPT users

Unlikely Resolves by Oct 31, 2027

Discussed by: OpenAI's product documentation, which currently excludes the Chat tab

OpenAI drops the individual Research approval requirement and makes the model available in the standard ChatGPT interface with the stricter refusals used by GPT-6 Astra. That would reverse the central design choice of this release: fewer refusals in exchange for vetted access.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

February 1975

Asilomar conference on recombinant DNA (1975)

About 140 molecular biologists met at Asilomar, California, to set voluntary containment rules for recombinant DNA research. The meeting followed a voluntary moratorium on certain experiments and preceded any commercial use of the technology.

Then

The group produced safety guidelines that shaped the U.S. National Institutes of Health's 1976 recombinant DNA rules.

Now

Asilomar became a template for scientists setting containment rules before regulators acted.

Why this matters now

GPT-Rosalind-6 is another case of experts deciding who gets access to a powerful biological tool before broad release.

February-November 2019

OpenAI's staged release of GPT-2 (2019)

OpenAI announced GPT-2 in February 2019 and initially withheld the full model, citing concern about misuse. It released smaller versions in stages through November 2019.

Then

Researchers could study smaller versions while OpenAI monitored for misuse.

Now

Staged release became a common, if contested, practice for frontier AI models.

Why this matters now

Rosalind's trusted-access model is a more formal version of the same tradeoff: capability withheld from the public to reduce misuse risk.

2011-2013

H5N1 gain-of-function controversy (2011-2013)

Two research teams showed that H5N1 bird flu could spread among ferrets after laboratory modification. The U.S. National Science Advisory Board for Biosecurity recommended publishing with redactions, and a voluntary moratorium followed.

Then

The episode produced the U.S. dual-use research of concern policy in 2012 and a 2013 resumption with conditions.

Now

It set the precedent that some pathogen research needs institutional review before publication or wider sharing.

Why this matters now

The debate over which biological work needs review, and who decides, is the same debate now applied to model access.