Claude Science: Anthropic's AI Workbench Built for Researchers, Not Just Chatting About Research

Anthropic announced Claude Science at a San Francisco launch event on July 1, 2026, positioning it as the scientific equivalent of Claude Code: a purpose-built application for researchers, not a general assistant that happens to know some biology.

CEO Dario Amodei described it as “our most significant expansion” of Anthropic’s life sciences work. The pitch is straightforward. Scientific research today is fragmented across dozens of databases, each with its own schema and query language, plus Jupyter notebooks, R sessions, HPC cluster terminals, PubMed, structural biology viewers, and more. Claude Science pulls those pieces into a single environment where researchers can run the full arc of a project without switching contexts every twenty minutes.

What Claude Science Actually Does

The workbench gives researchers access to a coordinating agent backed by over 60 prebuilt skills and connectors, pre-configured for genomics, proteomics, single-cell analysis, structural biology, cheminformatics, and related fields. It connects natively to databases like UniProt, PDB, Ensembl, ClinVar, ChEMBL, and GEO, and integrates with NVIDIA’s BioNeMo Agent Toolkit, giving access to models including Evo 2 for genomics and Boltz-2 and OpenFold3 for protein structure prediction.

One point Anthropic was clear about: this is not a new or more capable AI model. Claude Science runs on existing Claude models, including Opus 4.8, without any special biological training or privileged access. The improvement is at the application layer, not the model layer.

The agent architecture is worth understanding. A primary coordinating agent manages the overall project, spawning specialist sub-agents for specific tasks or routing work to custom agents that researchers have built themselves. Before any output reaches publication, a dedicated reviewer agent independently audits citations and calculations. That reviewer catches and corrects errors rather than just flagging them, which is a meaningful distinction for reproducibility.

The Reproducibility Problem, Addressed Directly

Any figure Claude Science produces comes bundled with a reproducibility package: the underlying code, the computational environment, a plain-language methodology explanation, and the complete message history. Other researchers can verify and replicate the work from that package.

You can also ask Claude Science to edit figures in plain language. Tell it to remove gridlines or switch an axis to log scale, and it edits the code itself. Sessions can be forked at any point to test two analytical approaches side by side without losing the original thread.

Compute Without the Friction

Compute management is handled through the workbench rather than requiring researchers to wrangle cluster jobs separately. Claude Science drafts a job plan, asks for confirmation before committing resources, and submits to whatever infrastructure the lab already uses, whether that is an HPC cluster over SSH or Modal for on-demand cloud compute. It scales from a single GPU to hundreds as the job requires.

Because the software runs on the lab’s own infrastructure, large or sensitive datasets do not need to leave existing systems. That matters particularly in clinical and pharmaceutical contexts, though it comes with a caveat: Claude Science is not HIPAA compliant during beta, so protected health information should not be used with it.

Early Results From Real Labs

Three early users offered concrete numbers at the launch.

Jérôme Lecoq at the Allen Institute used Claude Science to build an automated pipeline for scientific literature reviews. His team can now produce reviews exceeding 100 pages, work that previously stretched across up to two years.

At the UCSF Brain Tumor Center, epidemiologist Stephen Francis reported that comprehensive germline analysis of glioma that once took his group substantial time can now be completed in roughly a tenth of the time, with independently validated results.

Biotech firm Manifold Bio used Claude Science end-to-end to nominate drug targets for experiments, something it said a general coding assistant could not have done because the task required gathering the right data and applying judgment informed by the company’s own prior programs.

What This Means for You

If you work in a research environment, the practical question is whether Claude Science is ready to sit in your actual workflow or whether it remains a demo-friendly prototype.

The honest answer is: it depends on your setup. Claude Science currently supports macOS 13 and above, and Linux x64. Windows is not available at launch. It is released in beta for Claude Pro, Max, Team, and Enterprise users. Team and Enterprise admins will need to enable it for their organisations.

Pricing is not separate. Usage counts against your existing Claude plan limits. Anthropic has introduced a discounted Team plan specifically for active scientific labs at academic institutions and nonprofit research organisations, which is worth checking if your lab is on a tight budget.

For researchers who want to push further, Anthropic is funding up to 50 Claude Science AI for Science projects, each with up to $30,000 in credits. Modal is adding up to $2,000 in compute for selected projects. The focus is on biology and biomedical research, and applications are open through July 15, 2026.

The Broader Picture

Anthropic is not alone in this space. Google launched Gemini for Science at I/O in May, covering more than 30 life science databases. OpenAI released GPT-Rosalind in April, aimed at accelerating research and drug discovery. The timing of Claude Science, with its emphasis on reproducibility and compute integration, looks like a deliberate attempt to differentiate on research rigour rather than just breadth of data access.

One notable signal of confidence in Anthropic’s scientific direction: Nobel Prize winner John Jumper, who co-created AlphaFold at DeepMind, announced this month that he is joining Anthropic. Anthropic is also pursuing its own pre-clinical drug programs for neglected diseases, using Claude Science internally, both to contribute to science and to understand in practice how the tool holds up under real research conditions.

Claude Science is available now in beta at claude.com/science.