# GPT-6 Rosalind Research: API Pricing, Access and What Changed

OpenAI's new life-sciences release brings Astra's model weights into a restricted research deployment, with higher token prices and a separate approval process.

By Paras Madan | 2026-10-31


OpenAI released GPT-6 Rosalind Research on October 30 for a limited set of approved life-sciences customers. It brings GPT-6 capabilities to Work, Codex and the API, rather than the Chat tab. Its listed API input and output prices are three times those of the previous Rosalind Research model.

The important distinction is how OpenAI is packaging those capabilities. This is a restricted scientific deployment, with eligibility and oversight requirements that sit alongside the model itself.

## What changed from GPT-5.5 Rosalind Research

OpenAI's [model release notes](https://help.openai.com/en/articles/9624314-model-release-notes) say the API identifier is `gpt-6-rosalind`. GPT-5.5 Rosalind Research remains available, while the separate Rosalind Discovery tier stays on 5.5.

The [GPT-Rosalind-6 system card](https://deploymentsafety.openai.com/gpt-rosalind-6) says the release uses the same weights as GPT-6 Astra, without additional training data or post-training for this release. OpenAI describes a research deployment with different biological assistance boundaries and additional access controls.

**The product difference includes who can use it and under what conditions.** A specialist AI product does not always mean a separately trained base model. For a research organization, the permitted workflow, security requirements and review process can be just as important as the underlying model family.

## GPT-6 Rosalind API pricing

OpenAI's [life-sciences pricing table](https://developers.openai.com/api/docs/pricing) lists the following rates per million tokens:

| API model | Input | Output | Cached input |
| --- | --- | --- | --- |
| `gpt-6-rosalind` | $15 | $75 | $1.50 |
| `gpt-rosalind-research` | $5 | $25 | $0.50 |
| `gpt-rosalind-discovery` | $5 | $25 | $0.50 |

The new model also lists cache-write rates of $18.75 for 30-minute retention and $37.50 for 12-hour retention, per million tokens. Teams should check the current pricing page and applicable configuration when estimating a workload.

For a simple illustration, one million uncached input tokens and 100,000 output tokens would cost $22.50 at the listed GPT-6 Rosalind rates, compared with $7.50 for the older Research model. That calculation excludes tools, cache writes and other applicable charges. It is a token-cost comparison, not a prediction of the cost to complete a research task.

The practical question is whether better results reduce repeated work or expert review enough to justify the higher rates. A more capable model can still be the wrong default for straightforward tasks that an existing workflow handles reliably.

## Who can access GPT-6 Rosalind Research

The [access documentation](https://help.openai.com/en/articles/20001193-gpt-rosalind-for-life-sciences-research) distinguishes organizational approval from individual authorization. Research approval is required; Discovery approval does not automatically qualify an organization for Research. API access requires separate approval where applicable.

Enterprise workspaces begin with a two-week Admin Preview, with the model off by default. Administrators can enable it for approved Research users and configure access after the preview. Existing Research approval does not bypass that administrative step.

For teams planning an evaluation, this makes access a project dependency. Having an account, a paid subscription or approval for a different Rosalind tier should not be treated as confirmation that the new model is available to every researcher.

## What the safety evidence does and does not establish

OpenAI classifies the release as High capability in the Biological and Chemical domain, below its Critical threshold. Its system card reports mixed results against GPT-Rosalind-5.5 on the Critical-capability evaluations, with none reaching their indicative thresholds.

These are OpenAI's evaluations and framework classifications. They do not establish that the model will improve a particular laboratory's results or that its output can be accepted without review. The card also requires qualified human review before outputs inform experiments, publications or consequential decisions.

A useful evaluation should therefore begin with a bounded research workflow and a clear review standard. Record whether the model preserves evidence, identifies uncertainty and reduces the time experts spend correcting its output. That gives the higher token price something meaningful to be compared against.

The next signal to watch is independently reviewed evidence of better research decisions. More capable reasoning is valuable when scientists can trace, challenge and test what it produces.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn complex AI research and product launches into clear stories without overstating the evidence.

[Book a 30-minute call](https://cal.com/paras-madan/30min).