A biologist at Dyno Therapeutics submitted an RNA sequence prediction task to an AI model last month. The model's best ten submissions landed above the 95th percentile of human experts. These weren't undergrads or fresh postdocs. These were the people who design RNA for a living.

OpenAI named the model Rosalind, after Rosalind Franklin — the chemist whose X-ray images cracked DNA before Watson and Crick walked off with the Nobel. That's either a beautiful tribute or a cosmic joke, depending on how you feel about restricted-access AI models being named after a woman who was denied access to credit.

GPT-Rosalind launched on April 16, 2026. It's OpenAI's first reasoning model built specifically for life sciences. It scored 0.751 on BixBench, the bioinformatics benchmark from Edison Scientific — the highest score among models with published results. On Dyno's RNA prediction task, built with unpublished data to prevent contamination, it hit the 95th percentile for prediction and the 84th for generation.

The training wasn't generic. OpenAI says the model mastered 50 specific biological workflows during its training run — protein folding, pathway analysis, sequence-to-function prediction, the stuff PhDs spend years learning. Alongside the model, OpenAI released a Life Sciences plugin for Codex that connects to more than 50 scientific tools and biological databases. A wet lab in code.

The Access List

Here's where it gets uncomfortable. GPT-Rosalind is not for you. It's not for me. It's not for most of the world's working biologists.

The "research preview" is restricted to a "trusted-access programme" — limited to qualified enterprise customers in the United States. The launch partners are Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific. OpenAI says access is reserved for "organizations conducting legitimate scientific research with clear public benefits."

Translation: OpenAI is now in the business of deciding which scientific research is legitimate and which isn't. A grad student at a Brazilian university designing a tuberculosis RNA vaccine? Not on the list. A Kenyan biotech trying to cut sickle-cell trial costs? Not on the list. A pharma giant with $25 billion in revenue? Welcome, here's your key.

The Real Play

The claim is safety. The effect is moat-building. OpenAI just gave four of the most well-resourced biotech entities on Earth a tool that performs at the 95th percentile of their own experts. Everyone else waits.

This matters because drug discovery has always been gated by compute and data. Eli Lilly just brought online a 1,016-GPU NVIDIA Blackwell SuperPOD — 9,000 petaflops of AI to screen billions of molecules. Now OpenAI gives big pharma a reasoning engine to direct that compute. The gap between who can design drugs and who cannot just doubled overnight.

Google DeepMind did this first with AlphaFold — but AlphaFold eventually went public. Anyone can query it. GPT-Rosalind goes the other way. Private preview. Enterprise approval. American only.

My Opinion

I'll be blunt: calling a gatekept model "Rosalind" is the tech industry at its most tone-deaf. Franklin's entire story is about being locked out of the club while men took her data. Naming an exclusive access program after her is the kind of thing that happens when your marketing team reads Wikipedia but not the book.

The capability is real. The 95th-percentile claim is substantiated with unpublished data, which is the right way to run a benchmark. OpenAI deserves credit for building a model that actually beats the humans it's supposed to augment. This is a genuine capability jump in AI for biology.

Here's what bugs me. OpenAI is not a regulator. It's not the FDA. It's not a bioethics board. But it just appointed itself the gatekeeper of who gets to do "legitimate" drug discovery in the AI era, and the answer is "four American companies with seven-figure compute budgets." Drug discovery cannot be the privilege of the four biggest clients of the most valuable AI company in America. Before the end of this year, either OpenAI opens this up meaningfully, or an open-source competitor out of Shanghai or Paris will make the exclusivity irrelevant. I'm betting on the second one.

Author: Yahor Kamarou (Mark) / www.humai.blog / 18 Apr 2026