GPT-Rosalind Levels Up: Beating GPT-5.5 With 31% Fewer Tokens

Hey everyone, it's Shii-chan! Today I found a fun one where biology and AI meet: OpenAI's life-sciences model, GPT-Rosalind, just leveled up in a big way.

OpenAI News openai.com

What was announced?

This comes from OpenAI's News. The GPT-Rosalind series, purpose-built for life sciences research at enterprise scale, is getting a new model update.

The headline is that it fuses GPT-5.5's agentic coding and tool use with stronger intelligence in core drug-discovery domains like medicinal chemistry and genomics, and it also improves across broader analysis, design, and experimental workflows. It's available now as a research preview to eligible organizations worldwide through OpenAI's trusted-access deployment structure.

Why it matters

Progress in the life sciences depends on synthesizing data and evidence across scales and modalities: molecules, genes, pathways, and living systems. That's demanding even for human experts.

A model that can help across all of that could meaningfully change the speed and quality of research. And this update improves accuracy while spending fewer tokens, so it got smarter and more efficient at once.

What changes

Researchers can connect evidence that used to live in separate places, like literature, genomics, transcriptomics, sequence, structure, and experimental results, and move from data to research decisions more smoothly.

In OpenAI's evaluations, the update posts broad gains on expert research tasks, complex medicinal chemistry queries, quantitative biology, and even wet lab troubleshooting.

Dive Deep

First, the benchmarks are solid. OpenAI built LifeSciBench, an externally expert-judged benchmark focused on the foundations of life sciences research. It draws tasks from six workflow areas (evidence handling; analysis; design and optimization; scientific reasoning; validation and operations; and translation and scientific communication) and evaluates them end-to-end instead of one skill in isolation.

Here are the concrete numbers.

• On MedChemBench, GPT-Rosalind scores 27.5% vs. GPT-5.5's 25.1%, while using 7.2% fewer tokens.
• On GeneBench, a long-horizon agentic evaluation for genomics and quantitative biology, it reaches 21.6% vs. 20.4% accuracy while using 31% fewer tokens.
• On LabWorkBench, which uses real wet lab protocols, it scores 63.2% vs. 55.8%, using 5.3% fewer tokens. The data is proprietary, so it stays uncontaminated.

There's also a way to go from reasoning to execution. The Life Sciences Research and Life Sciences NGS Analysis plugins bring sourced evidence retrieval, biological interpretation, and bioinformatics execution into one workspace. The nice part: all users can access both plugins through Codex. Only qualified enterprise users, though, can power them with GPT-Rosalind itself.

Codex also added interactive viewers for biologically native file types, like sequence, alignment, and structure viewers, so scientists stay close to the evidence while the model reasons across a workflow.

Access is for organizations doing legitimate research with clear public benefit, strong governance and safety oversight, and controlled access with enterprise-grade security. Novo Nordisk is named as a new partner, and OpenAI is also offering a managed workspace for organizations without an Enterprise account. If your org wants in, start from the access request form.

Wrap-up

• A new GPT-Rosalind update that pairs GPT-5.5's agentic abilities with stronger drug-discovery intelligence.
• It beats GPT-5.5 on MedChemBench, GeneBench, and LabWorkBench while spending fewer tokens (31% fewer on genomics).
• The Life Sciences Research and NGS Analysis plugins are available to all users through Codex.
• It ships as a research preview via the trusted-access structure, with plans to extend to high-public-benefit areas like Rosalind Biodefense.

If you're a researcher rolling up your sleeves in drug discovery, medicine, or genomics, this is the one that hits hardest. The model itself is still limited to select organizations, but the plugins and viewers are open to Codex users, so if bio-meets-AI excites you, go take a look!