OpenAI brings GPT-5.5 agent smarts to life sciences with GPT-Rosalind

Pharmaceutical research usually hits one of two walls: finding the right lead molecule is like sifting through an ocean of sand, or a wet lab experiment fails and nobody knows which reagent went sideways.

OpenAI just rolled out an update to GPT-Rosalind, its purpose-built model series for enterprise life sciences. It fuses GPT-5.5's agentic coding and tool calling with deep domain intelligence in medicinal chemistry and genomics.

Access is managed through OpenAI's trusted-access framework for qualified organizations, with an update noting the model exited research preview on September 11, ahead of published pricing taking effect on October 5. Novo Nordisk is already using the model to help researchers sift through datasets and pressure-test hypotheses.

Fixing wet labs while burning fewer tokens

Instead of relying on generic chat benchmarks, OpenAI brought in evaluations tailored to actual benchwork.

LabWorkBench tests whether the model can link operational changes in real wet lab protocols to experimental outcomes, helping teams troubleshoot why an assay failed. Using proprietary, uncontaminated data, GPT-Rosalind hit 63.2% accuracy versus GPT-5.5's 55.8%, while consuming 5.3% fewer tokens.

On MedChemBench, which evaluates retrosynthesis, toxicity, and lead optimization, the model scored 27.5% compared to 25.1% for GPT-5.5, saving 7.2% in tokens.

For long-horizon genomics tasks on GeneBench, covering workflows like spatial transcriptomics and proteomics, GPT-Rosalind improved accuracy to 21.6% from 20.4%, but slashed token usage by 31%.

Across LifeSciBench, an expert-judged suite spanning six workflow stages from evidence handling to validation, GPT-Rosalind also topped both GPT-5.5 and Grok.

Native viewers land inside Codex

Biologists cannot spend their days copy-pasting FASTQ files into a plain chat box.

To bridge the interface gap, OpenAI rolled out two Codex plugins: Life Sciences Research and Life Sciences NGS Analysis. Any user can run them in Codex, but qualified enterprise teams can power them directly with GPT-Rosalind.

Codex now includes interactive visualizers for native biological data, including sequence viewers, alignment tools, and molecular structure visualizers. Researchers can ask questions and run analyses directly against the open file in view without switching windows.

For eligible research teams without an existing Enterprise contract, OpenAI is also offering a managed workspace option.

Why it matters

Life sciences is where general-purpose models typically fall apart. A generic chatbot sounds convincing until it hallucinates an impossible synthesis route or misinterprets an ADME profile.

We think pairing GPT-5.5's tool execution directly with domain-specific biological data is a sharp move. For research teams, the immediate upside isn't replacing medicinal chemists; it is automating the administrative grunt work between reading papers, running bioinformatics pipelines, and inspecting molecular structures.

That said, the barriers to entry remain deliberate and high. OpenAI is keeping GPT-Rosalind behind strict safety reviews and enterprise security checks, driven by biodefense and dual-use oversight. Independent researchers and casual experimenters will not be touching this anytime soon.

What this really signals is a shift in the AI biology landscape. The battleground is no longer just predicting folded protein structures; it is about building autonomous agents capable of troubleshooting messy, end-to-end wet lab reality.