Claude Science: Anthropic's AI Workbench for Researchers Explained
Anthropic Just Launched Its Most Ambitious Product Yet
On June 30, 2026, Anthropic released Claude Science, a product the company describes as an AI workbench for scientists. If Claude Code transformed how developers write software, Claude Science aims to do the same for how researchers conduct experiments, analyze data, and publish findings.
This is not a new model. Claude Science runs on the same Claude models already available to everyone, including Claude Opus 4.8. What makes it different is the environment built around those models: a single workspace that integrates more than 60 scientific databases, specialist AI agents, native visualization of molecular structures, and flexible compute management that scales from a laptop to hundreds of GPUs.
For anyone working in computational biology, drug discovery, genomics, or any data-heavy scientific field, this is a significant development. Here is everything you need to know about what Claude Science does, how it works, and why it matters.
The Problem Claude Science Solves
Scientific research in 2026 remains surprisingly fragmented. A typical computational biologist might start their day querying UniProt for protein data, switch to Ensembl for genomic annotations, pull expression data from GEO, check drug-target interactions on ChEMBL, review literature on PubMed, write analysis code in Jupyter, and submit compute jobs to an HPC cluster. Each of these systems has its own interface, schema, query language, and authentication.
This constant context-switching is not just inconvenient. It slows research down, introduces errors when data is manually transferred between tools, and makes reproducibility difficult. When a reviewer asks how a specific figure was generated six months after publication, reconstructing the exact pipeline from memory and scattered notebooks is often painful.
Claude Science addresses this by bringing all of these tools, databases, and compute resources into a single conversational interface. You describe what you want to investigate in plain language, and the system handles the orchestration across all the underlying services.
How Claude Science Actually Works
At its core, Claude Science uses a multi-agent architecture. When you start a session, you interact with a generalist coordinating agent that has access to over 60 curated skills and connectors. These are pre-configured for major scientific domains including genomics, single-cell analysis, proteomics, structural biology, and cheminformatics.
The coordinating agent can spin up specialist sub-agents for specific tasks. If you ask a question that requires pulling data from multiple sources, separate agents handle each query in parallel and the coordinator synthesizes the results. A dedicated reviewer agent then checks citations, validates calculations, and flags any inconsistencies before presenting the final output.
This architecture means Claude Science can handle complex, multi-step research workflows that would normally require hours of manual work. You can ask it to identify potential drug targets for a specific disease, and it will autonomously search relevant databases, assess surface expression and safety profiles, cross-reference the literature, and present ranked candidates with supporting evidence.
Database and Domain Coverage
The 60-plus databases and tools that Claude Science connects to span the major life sciences domains. In biology alone, this includes UniProt for protein sequences and annotations, PDB for protein structures, Ensembl for genomic data, Reactome for pathway information, ClinVar for variant-disease relationships, ChEMBL for bioactivity data, and GEO for gene expression datasets.
Claude Science also integrates with NVIDIA's BioNeMo Agent Toolkit, giving it native access to specialized life sciences models and libraries including Evo 2, Boltz-2, and OpenFold3. This means you can run protein structure predictions, analyze evolutionary sequences, and perform molecular modeling directly within the same environment where you are analyzing your data.
Visualization and Reproducibility
One of the more practical features is how Claude Science handles scientific visualization. It natively renders 3D protein structures, genome browser tracks, chemical structures, and publication-quality figures. You can interact with these visualizations in conversation, asking the agent to adjust axes, change color schemes, remove gridlines, or switch to log scale, and it edits its own code accordingly.
Every figure Claude Science generates comes with the exact code and environment that produced it, a plain-language description of the methodology, and the full message history. This built-in audit trail makes reproducibility straightforward. If you need to revisit an analysis months later, the complete provenance is preserved.
Compute Management
Large-scale scientific analyses often require significant compute resources. Folding a protein, running a genomics pipeline over a massive dataset, or training a model on experimental data can demand anything from a single GPU to a full cluster.
Claude Science handles this compute orchestration for you. It drafts a plan, requests approval before accessing new resources, and then writes and submits jobs to whatever infrastructure your lab already uses. This could be your own HPC cluster accessed over SSH, or a Modal account for on-demand cloud compute. It scales automatically from a single GPU to hundreds as the analysis requires.
Because the agents work within a running session that holds context in memory, even massive datasets only need to be loaded once. You can also fork a session at any point to compare two different analytical approaches without losing the original thread of work.
Data Privacy and On-Premises Deployment
For many research institutions and pharmaceutical companies, data privacy is non-negotiable. Clinical trial data, proprietary compound libraries, and patient genomic data often cannot leave institutional servers under any circumstances.
Claude Science addresses this directly. The platform can run on a lab's own infrastructure, whether that is a local workstation, a Linux server, or an HPC login node. Sensitive datasets never have to leave the systems they are already on. Only the context needed for each specific step of the analysis is sent to Claude's models. This architecture means that institutions handling proprietary or regulated research data can use Claude Science without compromising their data governance requirements.
This is a meaningful differentiator compared to some competing platforms that require all data to pass through external cloud services.
What Researchers Are Already Doing With It
Several research groups have been using Claude Science during its beta period, and the early results suggest meaningful productivity gains.
Multi-Agent Literature Reviews
Jerome Lecoq, a neuroscientist at the Allen Institute, used Claude Science to build what he calls a computational review template. This is a pipeline of about 20 custom skills where sub-agents read through thousands of papers, extract central claims and key quantitative findings, and store them in an evidence database. The system then constructs a narrative arc and writes the review section by section, with dedicated agents generating cross-study figures directly from the evidence.
A key innovation in Lecoq's workflow is the use of actor-critic pairs: one agent creates content while a separate reviewer agent evaluates it for accuracy and citation fidelity. Before Claude Science, reviews of this scope could take his team up to two years. He now has about 10 completed reviews, many exceeding 100 pages, with citations verified by reviewer agents.
Drug Target Identification
Manifold Bio, which designs tissue-targeting medicines, used Claude Science to nominate targets for its latest experiments. For each tissue and target, Claude Science assessed surface expression, trafficking, and safety, ranking candidates against criteria that Manifold has developed from its own proprietary data. According to Manifold, what distinguished Claude Science from a general coding assistant was its ability to handle this process end-to-end, gathering the right data and applying judgment with the context of past programs built in.
Brain Tumor Research
Stephen Francis, an associate professor and epidemiologist at the UCSF Brain Tumor Center, has used Claude Science for studies on the molecular epidemiology of glioma. His lab investigates how thousands of small-effect germline variants combine to shape individual susceptibility to brain tumors. Francis reported that Claude Science has enabled comprehensive germline workups across multiple approaches in roughly one-tenth the time it previously required. His group independently validated the results, confirming that the analyses were both rapid and robust.
Claude Science vs GPT-Rosalind
The most obvious comparison is with OpenAI's GPT-Rosalind, which launched in April 2026 as a specialized model for life sciences research. The two products take fundamentally different approaches.
GPT-Rosalind is a purpose-built model trained specifically for biological reasoning. It was designed from the ground up for scientific tasks and connects to more than 50 public multi-omics databases, including AlphaFold for protein structure lookup, Bgee for gene expression data, and BindingDB for ligand-target affinity. It is available through a trusted access program with partners including Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific.
Claude Science, by contrast, is not a new model at all. It wraps Anthropic's existing Claude models in workflow tooling, database connectors, and compute management. Anthropic is betting that the models they already have are capable enough, and that what scientists actually need is better integration and infrastructure around those models.
The other major difference is deployment flexibility. Claude Science can run entirely on local infrastructure, while GPT-Rosalind operates through OpenAI's cloud. For institutions with strict data governance requirements, this could be a deciding factor.
Both platforms connect to roughly similar numbers of scientific databases (60-plus for Claude Science, 50-plus for GPT-Rosalind), and both support multi-step research workflows. The competitive landscape will likely be shaped by which approach proves more effective in practice: a specialized model versus a general model with specialized tooling.
Availability, Pricing, and the Neglected Diseases Program
Claude Science is available in beta for Claude Pro, Max, Team, and Enterprise users on macOS and Linux. Team and Enterprise users will need their admin to enable it.
Anthropic is offering a discounted Team plan for active scientific labs at academic institutions and nonprofit research organizations, which could make the platform accessible to researchers who might otherwise find enterprise AI tooling prohibitively expensive.
The company is also supporting up to 50 Claude Science AI for Science projects with up to 30,000 dollars in credits per project. Modal is providing up to 2,000 dollars in additional compute credits for select projects. Applications are open through July 15, 2026, with awards announced by July 31 and projects running from September through December 2026.
Perhaps the most notable initiative is Anthropic's neglected diseases program. The company is launching an internal drug discovery effort focused on diseases that traditional pharmaceutical companies would not consider commercially attractive. Jonah Cool, Anthropic's head of life sciences partnerships, described these as areas outside the scope of what the traditional pharma and biotech landscape would target, but which nonetheless carry significant human burden. Major pharma partners including Novo Nordisk, AstraZeneca, and Eli Lilly are already using Claude in their own drug discovery workflows.
What This Means for Claude Users
Even if you are not a scientist, the launch of Claude Science signals a few important things about where Anthropic is heading.
First, it validates the multi-agent architecture pattern that is becoming increasingly central to how Claude operates. The coordinating agent, specialist sub-agents, and reviewer agent pattern in Claude Science is conceptually similar to what dynamic workflows and parallel sub-agents do in Claude Code. Anthropic is clearly investing heavily in this paradigm.
Second, the emphasis on reproducibility and audit trails suggests that Anthropic is serious about making AI outputs verifiable and trustworthy, not just fast. This focus on provenance and transparency is likely to influence how other Claude products evolve.
Third, the on-premises deployment option indicates that Anthropic recognizes data sovereignty as a first-class concern, not an afterthought. As more enterprises adopt AI tools for sensitive workloads, this flexibility will matter beyond just the life sciences.
For Claude power users who want to stay on top of their usage as these new products expand what Claude can do, tools like Gaugr can help you monitor your consumption across models and features in real-time.