Summary:

GPT-6 Rosalind Research is rolling out to approved organizations in Work, Codex and the API. Its developer-reported benchmark gains come with access, comparison and cost limits.

OpenAI began rolling out GPT-6 Rosalind Research on October 30 to a limited group of approved life-sciences research customers. The release brings a new option to Work, Codex and separately approved API organizations, rather than opening the model to everyone with a ChatGPT subscription. It is not available in the Chat tab. OpenAI’s release notes identify the API model as gpt-6-rosalind.

What changed

The system card calls the model GPT-Rosalind-6 and says it shares GPT-6 Astra’s weights, with no additional training data or post-training for this release. That distinction matters: this is a specialized research deployment, not evidence of a separately trained foundation model.

Three benchmarks, with important limits

OpenAI’s technical addendum, pages 3–4, reports the following results for Rosalind-6 versus rerun Rosalind-5.5:

  • MedChemBench v7: 49.6% versus 28.9%.
  • GeneBench Pro v13: 37.7% versus 16.7%.
  • LifeSciBench Gold: 61.3% versus 44.0%.

These are developer-reported results at the highest tested reasoning budget, using matched benchmark versions and problem sets. Metrics differ across rows; GeneBench uses a benchmark-specific configuration, and LifeSciBench excludes six tasks. The scores should not be averaged. Comparable Astra or rival-model results are not supplied for this trio, so the cover compares only the two Rosalind versions. This is not an independent leaderboard or clinical validation.

Approval, not just a subscription

OpenAI’s access guide separates Research and Discovery permissions. Discovery approval does not unlock Research; Discovery continues to use the existing 5.5 model. Existing customers can also retain GPT-5.5 Rosalind Research for established workflows.

Enterprise deployments begin with a two-week Admin Preview that is off by default. Administrators must enable eligible users. API access requires separate approval and is currently for internal research, not customer-facing products. The Rosalind Academic Plan does not include API access or credits.

What it costs

The API pricing page lists gpt-6-rosalind at $15 per million input tokens, $1.50 per million cached input tokens and $75 per million output tokens. Existing Rosalind Research is listed at $5 input and $25 output. Those are API rates, not a quote for an institution’s full deployment or ChatGPT usage.

What research teams should evaluate next

For an approved organization, the useful next question is whether the new deployment improves its own evidence-review and analysis workflows enough to justify the additional cost. A controlled comparison against existing workflows should examine factual accuracy, source traceability, human-review effort and total task cost. The published scores alone cannot establish those outcomes.

Cover: original PromptsZone scorecard rendered from OpenAI’s October 30 technical addendum, pages 3–4. Bold marks the higher score in each row; shading identifies the announced model. No source artwork was reproduced. Sources checked October 30, 2026. PromptsZone has not independently tested the model.