openai.com web signal

OpenAI Ships GPT-Rosalind-6 Biology Model to Trusted Users

6 sources tracking this story
OpenAI Healthcare Safety ai-business

TL;DR

  • GPT-6 Rosalind reuses GPT-6 Astra weights without separate training; the entire product differentiation comes from refusal configuration and institutional access controls.
  • The 1.5x price premium works out to $15/$75 per million input/output tokens versus $10/$50 for standard Astra access, charged entirely for governance configuration.
  • Rosalind scores 0.751 on BixBench Pass@1, the highest published result on this bioinformatics benchmark, outperforming GPT-5.4 on 6 of 11 LABBench2 tasks.

OpenAI on October 30 published a system card addendum for GPT-Rosalind-6, described as "a frontier reasoning model for biology, drug discovery, and translational medicine" that ships with the same weights as GPT-6 Astra but is reserved for trust-based access.

The gating is the point. OpenAI says the model is "designed to support legitimate biological research, including some dual-use topics, while still refusing requests for biological weaponization," and is tuned "to refuse or safely respond to requests that would continue to meaningfully enable severe biological weaponization." Non-government applicants undergo business verification and compliance screening; government applicants follow a separate trusted-access pathway for biosecurity, biosafety, public-health preparedness, and biodefense. Enterprise customers must apply "qualified human review before outputs inform experiments, publications, or decisions."

On OpenAI's internal benchmarks the capability jumps over GPT-Rosalind-5.5 are large. MedChemBench v7 moved from 28.9% to 49.6% across 99 problems. GeneBench Pro v13 rose from 16.7% to 37.7% across 129 problems. LifeSciBench Gold went from 44.0% to 61.3% across 94 problems. On Biorisk Knowledge consensus-at-32 the model scored 78.3% unadjusted and 85.0% refusal-adjusted, above the 80% indicative High threshold, with a 9.4% refusal rate.

The addendum rates the model High in Biological and Chemical, below the Critical threshold, and keeps the Critical cybersecurity rating from GPT-6 Astra under what the card calls a "more conservative cyber refusal boundary." "None of the four Critical capability evaluations met the indicative threshold," the card states, naming AAV capsid packaging prediction (developed by Dyno Therapeutics), SHP2 protein function prediction, a coronavirus-ACE2 cell-entry screen, and phage-plasmid co-evolution.

OpenAI notes that biosecurity red-teaming was conducted on GPT-Rosalind-5.5 before launch and is being repeated for GPT-Rosalind-6. The release lands amid a dense run of OpenAI coverage on our tracker, a day after the lab restricted teen weapons information and auto-ended violence chats in ChatGPT — two decisions that point in opposite directions on where OpenAI thinks the safety line sits for different audiences.

What others are reporting

Coverage cluster as of 24h after publish

  1. OpenAI Read →

    First-party canonical page. Establishes the dual-track model: commercial institutional vetting and a direct government pathway for public health and biodefense missions.

  2. Tech Insider Read →

    Provides benchmark data (0.751 BixBench Pass@1), expands named partner list to Thermo Fisher and Allen Institute, and frames Rosalind as direct competition to Microsoft Azure AI for Health.

    Not a chatbot, not an image generator, and not a coding copilot. It reasons over molecules, proteins, genes.
  3. i-SCOOP Read →

    Maps OpenAI's approval requirements against EU AI Act, GDPR, and NIS2, showing European institutions face dual compliance burden and treating the vetting process itself as the product sold.

    What OpenAI really sells here is a licence to ask harder questions, under watch.
  4. FourWeekMBA Read →

    Documents exact token pricing ($15/$75 vs $10/$50 per million) and confirms the premium is charged for governance configuration alone, with no new training or post-training involved.

    Rosalind has fewer refusals on dual-use biological queries but is only made available via trusted access.
  5. Data Studios Read →

    Confirms Rosalind is a deployment configuration of Astra rather than a separately trained model, centering the buyer question on refusal policy and access gate rather than capability uplift.