---
title: "OpenAI rolls out GPT-6 Rosalind Research with restricted access and new benchmark results"
description: "OpenAI releases GPT-6 Rosalind Research for approved organizations. Compare three reported benchmarks and understand access restrictions and API pricing."
url: "https://promptszone.com/openai-gpt-6-rosalind-research-release-benchmarks-access"
published: "2026-10-30T22:00:05Z"
categories: ["New Models"]
---

# OpenAI rolls out GPT-6 Rosalind Research with restricted access and new benchmark results

OpenAI’s October 30 Rosalind release brings GPT-6 to approved life-sciences workflows, with three benchmark comparisons, higher API prices and restricted access.

![Original scorecard comparing OpenAI-reported Rosalind-6 and rerun Rosalind-5.5 results on MedChemBench v7, GeneBench Pro v13 and LifeSciBench Gold. Higher scores are bold; methods and limitations are labeled.](https://promptszone.com/assets/uploads/602c1974642926f4c7364d5972fa61b1.png)

**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](https://help.openai.com/en/articles/9624314-model-release-notes) identify the API model as **gpt-6-rosalind**.

## What changed

The [system card](https://deploymentsafety.openai.com/gpt-rosalind-6) 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](https://deploymentsafety.openai.com/gpt-rosalind-6/gpt-rosalind-6.pdf), 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](https://help.openai.com/en/articles/20001193-gpt-rosalind-for-life-sciences-research) 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](https://developers.openai.com/api/docs/pricing) 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.