TL;DR: On August 28, 2026, OpenAI announced Rosalind Workbench — a research-preview environment for life-science work, available through the ChatGPT app and built into Codex. It runs on GPT-Rosalind, OpenAI's dedicated life-sciences model, and adds in-conversation molecular structure, sequence-alignment, and slide viewers plus a plan-first NGS Analysis Workbench. Advanced workflows are gated: general questions run in Explore mode, while Research mode needs a verified organization to request access, with individual access "coming soon."
The framing OpenAI uses is "every scientist to be their own research team" — a collaborative co-scientist that drafts plans, orchestrates tools, and returns reviewable outputs while humans keep responsibility for validation.
What people are asking
| Question | Short answer |
|---|---|
| Is it available now? | Research preview only, through the ChatGPT app and Codex. Not a general release. |
| Do I need an org? | Research mode does — a verified org member requests access for the organization. Explore mode is lighter. |
| What model powers it? | GPT-Rosalind, OpenAI's life-sciences model, first seen in the LifeSciBench benchmark. |
| Does it run wet-lab experiments? | No. It plans assays, prices them, and assists — humans run and validate the bench work. |
| What's the closest comparison? | Anthropic's Claude Science workbench; on autonomy, DeepMind's Co-Scientist. |
| Any third-party integrations? | Coverage names Boltz Bio and NVIDIA BioNeMo. |
What Rosalind Workbench actually is
Rosalind Workbench is a central environment where life-science users bring their favorite scientific tools, explore specialized biology models, and define data-analysis workflows. OpenAI describes it as a guided experience rather than a raw API surface — the workbench proposes a plan, you approve it, and it coordinates the underlying tools.
It is built on GPT-Rosalind, which OpenAI positions as frontier reasoning plus specialized tool orchestration across medicinal chemistry, genomics, and wet-lab assistance. The stated direction is "teams of agents" spanning those domains — a multi-agent research org rather than a single chat model. That mirrors the multi-agent pattern Anthropic has been testing in work like its mind-viruses multi-agent research.
The leads named on the launch are Chris Hayduk (Life Sciences at OpenAI) and Joy Jiao.
In-conversation viewers
The most visible new capability is a set of specialized viewers that open inside the chat, so structural and sequence work happens where the reasoning happens:
- Molecular Structure Viewer — for example, opening the biological assembly of PDB 5LF3, the human 20S proteasome bound to bortezomib. In OpenAI's GLP1R bound to semaglutide example, the model interacts with the viewer, explains structural features from the PDB file, and creates a movie of the structure.
- Biological Sequence and Alignment Viewer — for inspecting sequences and alignments in place.
- Slide Viewer — for pathology and presentation material.
This is the same "put the domain tool in the workspace" idea behind Claude Science, which wires PubMed, Jupyter, R, and structural-biology tooling into one auditable environment. The difference is packaging: Rosalind Workbench keeps everything inside ChatGPT and Codex.
Starter tasks
Rosalind Workbench ships with starter tasks across six areas: protein design, small-molecule design, safety and developability, structure and sequence, genomics and pathology, and experimental validation. Concrete examples from the announcement:
- Rank PD-L1 nanobody candidates, then plan and price binding assays for five nanobodies.
- Dock imatinib into ABL1 and inspect the five leading poses.
- Run a dexamethasone-treated airway RNA-seq task in the NGS Analysis Workbench.
The pattern is consistent: the model does the ranking, docking, or analysis, and then produces a concrete experimental plan with costs for a human to approve.
The NGS Analysis Workbench
The Rosalind NGS Analysis Workbench is a sequencing pipeline that takes FASTQ inputs, runs quality control (FastQC), and handles bulk RNA-seq or single-cell analysis. Its defining behavior is that it prepares a plan for researcher approval first, then coordinates the tools and returns traceable, reviewable outputs.
That plan-first, traceable-output design is the same accountability move we've seen elsewhere in scientific AI tooling — the GeneBench Pro work and Claude Science's reviewer agent both lean on making every step inspectable rather than trusting a single generated answer. For genomics specifically, it's a more product-shaped take on what benchmarks like GeneBench were measuring.
Explore mode vs Research mode
| Mode | For | Access |
|---|---|---|
| Explore mode | General scientific questions, using available ChatGPT models | Lighter — open in the research preview |
| Research mode | Complex biology, advanced workflows | Verified org member requests access on behalf of the organization; individual access "coming soon" |
If you are an individual researcher without an org sponsor, Explore mode is what you get today. This is a narrower on-ramp than ChatGPT for Academic Researchers, which handed free GPT-5.6 Sol Pro access to verified individuals — Rosalind Workbench's advanced tier is org-first.
Rosalind Workbench vs DeepMind Co-Scientist
The same day OpenAI shipped Rosalind Workbench, Google DeepMind published its execution-grounded Co-Scientist real-world paper. They are solving adjacent problems from opposite ends:
- Rosalind Workbench is the integrated workbench and tooling layer — viewers, an NGS pipeline, starter tasks, and plans a human drives and reviews. It is a co-pilot environment.
- Co-Scientist is an autonomous multi-agent research loop that includes proposing and running physical experiments, closer in spirit to Google's earlier AI Scientist / ScientistOne work.
Neither claims AI is now doing science unsupervised. Both keep humans on review and validation. The near-term question for a lab choosing between them is whether you want a better workspace or a more autonomous collaborator.
Honest limitations
- Research preview. No SLA, no pricing page, feature set will move.
- Advanced access is org-gated. Individual Research mode is "coming soon," not available.
- No wet-lab execution. It plans and prices assays; it does not run them. Independent wet-lab validation — the step that separated hype from result in Claude's protein-design campaign — is still on you.
- Claims are OpenAI's own. There is no third-party evaluation of Rosalind Workbench yet, and the broader field still has an evidence problem in AI drug discovery.
- Integrations with Boltz Bio and NVIDIA BioNeMo are mentioned in coverage rather than deeply documented at launch.
What builders and researchers should do
- If your org has an OpenAI enterprise relationship, have a verified member request Research mode access now — the preview is where feedback shapes the product.
- Otherwise, try Explore mode for structure and sequence questions and see whether the in-chat viewers change your workflow.
- If you already use Claude Science, run the same task in both and compare the plan quality and auditability, not just the final answer.
- Treat every generated experimental plan as a hypothesis to validate, not a result.
Related reading
- LifeSciBench: OpenAI's 750-task benchmark for GPT-Rosalind
- Claude Science: Anthropic's AI workbench for scientists
- Claude designed working protein binders for 14 of 15 targets
- ChatGPT for Academic Researchers: free GPT-5.6 Sol Pro access
- OpenAI GeneBench Pro: GPT-5.6 Sol on computational biology
- Can AI cure cancer? What the evidence actually shows
- Google's AI Scientist / ScientistOne: chain-of-evidence at ICML 2026
- AI drug discovery's evidence problem
Official source: OpenAI — Meet Rosalind Workbench (August 28, 2026)
Details reflect OpenAI's August 28, 2026 announcement. Rosalind Workbench is in research preview — access rules, modes, integrations, and capabilities are expected to change. Verify current access requirements with OpenAI before planning work around it.