Good evening, everyone. OpenAI put out a new model for biology labs yesterday, and most of us cannot use it. That is not a mistake in the launch. That is the whole point of it.

Quick take: GPT-6 Rosalind Research is OpenAI's new life sciences model, built on the same weights as its flagship GPT-6 Astra. It scores far higher than its predecessor on lab benchmarks, costs three times as much through the API, and only approved research organisations get in. A subscription alone does not open the door.

Benchmark (highest reasoning tested) — Rosalind 5.5 vs Rosalind 6
Medicinal chemistry: 28.9% → 49.6%
Genomics and quantitative biology: 16.7% → 37.7%
Applied life sciences: 44.0% → 61.3%

Same brain, different door

Here is the part that stopped me. OpenAI says Rosalind 6 uses the same underlying weights as GPT-6 Astra, with no extra training for this release. So this is not a brand new brain. It is the same brain, handed out with different rules.

The difference sits in how it handles biology questions, and in who gets to ask them. It refuses fewer legitimate biology questions that could also be misused, while still refusing requests for biological weapons. OpenAI rates it High for biological and chemical capability, below its Critical threshold, and Critical for cybersecurity because it inherits Astra's rating there.

That mix explains the gate. A general model that answers everyone can afford to be cautious. A specialist model that refuses less needs to know who is on the other side of the screen.

What it actually does in a lab

Rosalind is built for biology, drug discovery, and translational medicine. In plain terms, it helps researchers read papers, work through biological data, plan experiments, and think through medicinal chemistry.

The numbers in the table above are OpenAI's own, not independent tests, so take them as a claim, not a verdict. Still, the jumps are large. Medicinal chemistry went from 28.9% to 49.6%. Applied life sciences tasks set by working scientists went from 44.0% to 61.3%.

One honest catch comes from OpenAI itself. One biological data benchmark used a setup different from the production system. That does not kill the story, but it is a good reason to wait for outside labs to run their own checks before treating these scores as settled fact.

The price tag tells its own story

Through the API, Rosalind 6 costs $15 per million input tokens and $75 per million output tokens, with cached input at $1.50. Its predecessor cost $5 and $25. So the new version is three times the price of the old one.

You will not find it in the ordinary Chat tab. It shows up in ChatGPT Work and Codex for approved organisations, and API access needs a separate approval. Even inside an Enterprise workspace, it starts with a two-week admin preview that is off by default until an admin turns it on.

And approval for the older Discovery tier does not carry over. Each applicant faces a review of their research plans, security controls, and governance, and approved customers must limit use to authorised users with qualified human review of outputs.

Two gated models in one week

This is the second time in a few days I have seen this pattern. Google's Gemini 4 Carbon, announced on October 28, went first to a subset of cybersecurity defenders through its Fairwind Program, with wider access promised later and no date given.

Now OpenAI does the same for biology. The most capable specialist models are no longer launched to everyone at once. They are launched to a vetted list first.

I think that split is the real story here. General models keep getting cheaper and wider. Specialist models that touch biology or security are getting more gated, more expensive, and more checked. Both things can be true at once, and both tell you where the labs think the risk sits.

Bottom line: Rosalind 6 looks like a serious step for approved labs, and a closed door for everyone else. I would rather see the gate than a model that answers every biology question for anyone who asks. But the gate only works if the review behind it is real, and if outside researchers get to test these claims soon.

Would you trust a biology model more because it is gated like this, or does the gate just mean the rest of us never get to check its work?

Originally published on PeakD: https://peakd.com/hive-167922/@worthmind/openai-s-new-biology-model-is-

powerful-but-you-cannot-just-buy-it

Cover image: AI-generated