Model Drop

GPT-Rosalind

April 19, 2026

View on Substack

Today’s subject: GPT-Rosalind. This model is OpenAI’s first domain-specialized frontier model, shipped last week, is named after Rosalind Franklin, whose X-ray diffraction work made the double helix possible and whose credit got handed to Watson and Crick.

A fitting namesake for a model you can only use if OpenAI decides you qualify.

Benchmarks

Benchmark numbers worth noting:

• BixBench (real-world bioinformatics and data analysis): 0.751 pass rate, the strongest published score among models with public results

• LABBench2 (literature retrieval, database access, sequence manipulation, protocol design): outperformed GPT-5.4 on 6 of 11 tasks, with the biggest gains in CloningQA (end-to-end reagent design for molecular cloning)

• RNA sequence prediction on unpublished sequences: best-of-ten submissions ranked above the 95th percentile of human experts

• RNA sequence generation: 84th percentile of human experts

The positioning is that GPT-Rosalind reasons over molecules, proteins, genes, pathways, and disease biology, and uses scientific tools across multi-step workflows (literature review, sequence-to-function interpretation, experimental planning, data analysis) better than any general-purpose model.

What’s new

Scientific tool use as a first-class capability. The model was trained to query specialized databases, parse scientific literature, and drive computational tools inside a single interface.

Skepticism over sycophancy. OpenAI explicitly trained the model to reject weak drug targets and question fragile connections rather than validate the user’s framing. This is the first time OpenAI has shipped a model that’s supposed to disagree with you by default, an important differential given the domain.

Built-in biosecurity guardrails. Biological modeling has severe dual-use risks. OpenAI acknowledged this and built the access program around it: vetted institutions, secure environments, approved-user lists, compliance requirements.

Codex integration. When evaluated inside Codex, the model’s submissions hit the 95th percentile on prediction and 84th percentile on sequence generation. The Codex plugin is where the model is meant to actually work.

• Best fits: genomics analysis, protein engineering, biochemistry reasoning, literature synthesis, experimental protocol design, bioinformatics pipelines, drug target evaluation.

• Skip it for: anything outside life sciences, anything you need today (if you aren’t Moderna), anything that needs reproducibility across a public research community.

Impressions

Reception split along the fault line you’d expect. Pharma and institutional partners praised it. The broader AI and science community raised the access question before the benchmark question.

The positive

• Kevin Weil led the technical pitch. OpenAI’s CPO called it “our first frontier model built for scientific research across biology, drug discovery, and translational medicine,” with built-in knowledge of chemistry, protein engineering, and genomics.

• Partners endorsed it on the record. Moderna’s Stéphane Bancel called it “an important step in helping scientific teams use advanced AI to reason across complex biological evidence, data, and workflows.” The Allen Institute’s Andy Hickl said it makes manual steps like finding and aligning data “consistent and repeatable in an agentic workflow.”

The negative

• Sells access, not discovery. Implicator.ai’s read was that the actual product here is the gate, not the model: “OpenAI’s Biology Model Is Not a Lab Breakthrough. It Is an Access Strategy.” The argument is that ChatGPT, Codex, API, vetted customers, plugin, and biosecurity rules stitched together is the moat, and the benchmark numbers are the cover story.

Where it lands

GPT-Rosalind is the first serious domain-specialized frontier model from a major lab. If it works, expect GPT-Legal, GPT-Clinical, GPT-Materials, and five others by the end of the year.

The interesting part is the business model. OpenAI is testing whether vetted access to a specialized model inside Codex and ChatGPT is a product pharma will pay real money for, and whether biosecurity controls can ship as a feature rather than a compliance burden. If Moderna and Amgen renew after the preview at enterprise pricing, I think they’ll view it a success.

The namesake is the part that lingers. Rosalind Franklin’s data was used without her consent to build a breakthrough she wasn’t credited for. OpenAI built a model on decades of published and unpublished scientific work and gated access to the people most likely to already have institutional power. Worth sitting with.