GPT-Rosalind-6 Reuses GPT-6 Astra's Weights, Cuts Biology Refusals to 9.4% and Stays Behind a Gate
A.I. / news
GPT-Rosalind-6 Reuses GPT-6 Astra's Weights, Cuts Biology Refusals to 9.4% and Stays Behind a Gate
OpenAI's Oct. 30 system card addendum says no training was added, so the difference from Astra is deployment controls and who is allowed in.
OpenAI published a system card addendum on Oct. 30 for GPT-Rosalind-6, a life-sciences model that uses the same weights as GPT-6 Astra and is offered only to approved customers.
The addendum says: "We did not add training data or perform additional post-training for this release." It adds that Rosalind-6 has "additional deployment controls for its use in advanced life sciences research."
The refusal rate is the difference
The card summarises the trade in one line: "Rosalind has fewer refusals on dual-use biological queries but is only made available via trusted access." Its biorisk knowledge table shows what that means in numbers.
| Model | Score (cons@32) | Refusal rate | Refusal-adjusted |
|---|---|---|---|
| GPT-Rosalind-6 | 78.3% | 9.4% | 85.0% |
| GPT-Rosalind-5.5 | 79.2% | 1.8% | not listed |
| GPT-6 Astra | 63.3% | 29.2% | 90.0% |
The indicative High threshold on this test is 80%. Rosalind-6's raw score of 78.3% sits under it, and the refusal-adjusted 85.0% sits over it. The card says that after adjusting for refusals the model "exceeds the indicative High capability thresholds on three of the four evaluations." The page rates the model High in Biological and Chemical, below Critical.
Rosalind-6 refuses more often than its predecessor Rosalind-5.5, 9.4% against 1.8%, but about a third as often as Astra's 29.2%.
Capability scores are OpenAI's own
Against GPT-Rosalind-5.5 at the highest reasoning budget, the card reports 49.6% on MedChemBench v7 (99 problems) against 28.9%. On GeneBench Pro v13 (129 problems) it reports 37.7% against 16.7%, in a benchmark-specific configuration the card says is not the production setup. On LifeSciBench Gold (94 problems after excluding 6 flagged) it reports 61.3% against 44.0%.
Every one of these is vendor-supplied. The card names no outside party that reran them. Its Critical-level tests, including Dyno Therapeutics' AAV capsid packaging task with a 0.600 threshold, are described as not meeting the indicative thresholds.
In the page text The Terminal fetched, several results tables showed column headers without values. These include ProtocolQA Open-Ended, multi-select virology troubleshooting and the biosecurity safeguards table, so those figures are not reported here.
Who gets in
OpenAI says access is for approved customers whose life-sciences mission, use case, governance and ability to control access it has reviewed. Non-government applicants go through business verification and compliance screening. Government applicants use a separate pathway for biosecurity, biosafety, public-health and biodefense work. OpenAI says it can restrict or revoke access for misuse.
The requirements match those for GPT-Rosalind-5.5, "including later policy updates," and the card describes no broadening of eligibility. That model launched on April 16 as a research preview limited to qualified US enterprise customers, with Amgen, Moderna, Thermo Fisher Scientific and the Allen Institute as launch partners, The Next Web reported. It is named after Rosalind Franklin.
Implicator editor-in-chief Marcus Schuler argued at launch that the product is mainly governed access rather than proof that AI has made drugs. The card's refusal figures fit that reading, since the safeguard that changed is who can ask.
Cyber and alignment numbers
Rosalind-6 inherits Astra's cyber safeguards and its Critical cybersecurity rating. On a historical jailbreak regression with real-time blocking disabled, it scored 14.3% against Astra's 15.7%, where lower is better. Our report on Anthropic's exploit-chain research shows how vendors are now describing that class of capability.
On alignment tests the card lists mostly small gaps. Unintended Engagement, a specified misaligned action, rose to 2.5% from Astra's 0.0%. Blocker Deception fell to 9.5% from 12.9%. For how such agent evaluations behave in practice, see our Minecraft agent swarm report.
OpenAI's page gives no date for any wider release of Rosalind-6.
Sources
More in A.I.
- 01Architect Says Claude Opus 5.5 Drew a House Set Good Enough for a Permit After Two Redline RoundsStephen Gallagher's informal test of 13 models gave Opus 5.5 the most human feedback, and the drawings still missed a required second exit.
- 0250 GPT Agents in Minecraft Built a Ration Ledger, and 34 Died Doing ItY Combinator's Labs group and Olam Labs ran GPT-6.1 Sol and Claude Opus 5.5 swarms through Minecraft and Mindustry. GPT won both head-to-head games, but each result comes from one harness and one setup.
- 03China's Announcement No. 48 Puts 23 AI Technologies on a Draft Export ListExporters of listed technology must already register with MOFCOM. Comments on the draft catalogue close on November 15, and the sources disagree on whether the thresholds are published.
- 04Lean Creator Leo de Moura Says AI Makes Proofs Cheap. Here Is What the Checkers ShowOn a podcast released October 29, the AWS scientist who built Lean claimed AI-written proofs now run to 50 million lines. The claims check out only in part, and he names the two gaps himself.